Technical Documentation

{{branding.parent}} · {{product.nameUpper}}

Welcome to the employee page.

Everything you need to get up to speed on {{product.name}} and take it into a client conversation, in one place. Pick the path that matches what you are here to do; you can switch at any time with the tabs above.

Path 1 · Learn

I want to know more about {{product.name}}

Get up to speed on Blind Overspend, how cost per verified outcome is worked out, and who buys it, at your own pace.

A 3-minute introduction tour
Learning path and concept glossary
Searchable FAQ knowledge base

Start learning →

Path 2 · Sell

I want to sell {{product.name}}

Pitch to a CFO, CAIO, CIO or CTO, handle objections, position against named alternatives and explain the offer ladder.

Persona messaging and elevator pitch
Positioning, objections, offer ladder
The shared deal tracker

Start selling →

Path 3 · Demo

I want to demo {{product.name}}

Run the demo on the seeded demo tenants, with a script built on what the product does today.

Three demo tenants and what each shows
A 20-minute script and known pitfalls
Shared demo hub

Start demoing →

Path 4 · Build

I want to build with {{product.name}}

Get access to the app, see what is live and what is roadmap, and find the technical documentation.

Access, roles and who to contact
The release roadmap
Technical enablement portal

Start building →

LEARN {{product.nameUpper}} IN 3 MINUTES

The {{product.name}} introduction tour

Swipe through six slides: the problem, the solution, how it works, who it is for, what good looks like and how to start.

01 / 06 — The problem

Blind Overspend

Enterprises can report their total AI spend. They cannot report cost per use case, cost per outcome, or whether the human oversight protecting AI decisions is efficient. Blind Overspend is AI consumption that cannot be attributed to a business outcome. Agents add a second problem: one behaving exactly as designed can burn budget and take irreversible actions without any alarm.

No cost per outcome Agents that fail silently

02 / 06 — The solution

{{product.description}} It reads billing and usage from the client's AI providers, joins them to outcomes the client has signed off, and says plainly how much of the picture it can and cannot see.

{{product.category}} Outcome-linked, not just spend

03 / 06 — How it works

Four modules, one record

Govern: read-only provider connections, a cost baseline and a completeness score.
Attribute: signed outcome definitions, cost per verified outcome, and Verified Yield once a baseline exists.
Agentic: sessions authorised before launch, monitored in flight, audited afterwards.
Optimise: recommendations that are only actioned after a recorded human approval.

See the Platform page

04 / 06 — Who it is for

Buyers, users and sponsors

Economic buyer: the CFO, who has to defend AI spend at board review.
Delivery sponsor: the CAIO or CTO, who owns agents and the people to run governance.
Also in the room: the CIO, who asks about lock-in and data residency.
Best fit: enterprises spending materially on AI across more than one provider, especially regulated ones and those running agents.

05 / 06 — What good looks like

What the product has to prove

Coverage is honest: what share of spend it can see, per provider, with the gaps named.
The unit cost is real: cost per outcome from reconciled billing, inputs visible on demand.
Value is signed, not asserted: no signed outcome definition, no figure.
No client result has been cleared for internal sharing yet. Approved references are shared on request through the product team once they are cleared. Do not quote any.

06 / 06 — How to start

From first conversation to scale

1. Diagnostic: four weeks, fixed scope, a baseline and top five opportunities
2. Tier 1 to Tier 2: govern, then attribute to outcomes
3. Tier 3 and beyond: continuous optimisation, with the agentic add-on where agents run
Durations beyond the Diagnostic are scoped per engagement.

Request a session

Level 1 · Foundations

I can explain it

Understand Blind Overspend, cost per verified outcome and why coverage is shown honestly.

Do this: take the introduction tour above; read the concepts below; skim the FAQ; open the Govern page of the Northwind demo tenant and find the coverage badge next to the spend figure.

Suggested effort: about 1 hour · Outcome: explain Blind Overspend and CPVO to a colleague in two minutes

Level 2 · Practitioner

I can pitch and demo it

Run a discovery conversation with a CFO or CAIO and deliver the 20-minute demo.

Do this: work through the Sell tab (personas, objections, offer ladder); run the demo script on all three demo tenants; deliver it to a peer and take their questions.

Suggested effort: 1 to 2 days · Outcome: lead a first client call with a peer observing

Level 3 · Expert

I can design and deliver it

Scope an engagement, run checkpoints and understand how the platform is built and operated.

Do this: read the enablement portal (architecture, deployment, integrations); work through gates G1 to G7 on a demo tenant in the checkpoint console; read the design decision records and the runbook in the repository. No formal certification exists yet; ask the product team about a walkthrough with a capability owner.

Suggested effort: 1 to 2 weeks · Outcome: take a Diagnostic scope through to a signed-off baseline with a mentor

Who to ask: the {{product.name}} team at {{product.contactEmail}}. The product owner is the {{product.name}} product lead and enablement is run by the same team; the mailbox above routes to both. Support link, if configured: support channel.

The problem

Blind Overspend

AI consumption that cannot be attributed to a business outcome. It is the named enemy: the client can see the invoice but not what it bought.

The headline metric

CPVO: cost per verified outcome

The full cost of a workload over a period divided by the outcomes that were verified in that period. Cost includes tokens, supporting infrastructure, the oversight effort and any remediation. If the client has not signed what one outcome is, the product shows "not attributable" instead of an estimate. If there are too few verified outcomes, it shows the figure with a low-confidence label and the count.

Not on day one

Verified Yield

The value a workload created above the best available alternative, net of the full cost of the AI route. It needs a captured pre-AI baseline, so it belongs to Tier 2 onward. Never promise it in a first meeting.

Honest instrumentation

Coverage and completeness

Completeness is the share of a provider's actual spend the platform can account for. Coverage is how that shows up for the client: a percentage next to the spend figure, with the gaps named. A provider below 95% is flagged, not hidden. A gap is a next step, not an embarrassment.

One number to track

AYCI: AI Yield Confidence Index

A composite the CFO can follow period on period, always published with its four parts: coverage, attribution, verification and control. A single index without its parts is not defensible, and the parts are where the next conversation lives.

How IP is protected

Dual-mode

Every screen has an internal and a client projection, decided on the server. Clients get their own data, findings, provenance and coverage in full. They do not get the calculation parameters, thresholds or confidence model. Capgemini staff can switch to client mode to present; do that before you share your screen.

The journey

Gates G1 to G7

G1 scope and connection, G2 baseline acceptance, G3 governance framework, G4 outcome definition, G5 attribution acceptance, G6 optimisation authorisation, G7 sustained control review. Tier 1 covers G1 to G3, Tier 2 G1 to G5, Tier 3 G1 to G7. A separate gate, Gate A, authorises each agent session before it starts.

The operating model

Checkpoints

Each gate is a checkpoint in a console: one plain question, the platform's finding with a confidence signal, two to four bounded options, a visible timer, and an escalate button. Target is 20 minutes or less. L1 consultants run them, L2 experts take escalations, L3 capability owners calibrate the platform. Clients see "escalated for expert review".

Internal only: the agent monitor labels its states GREEN, AMBER and RED (AMBER pauses a session, RED stops it). With clients say "paused" and "stopped".

🔍

No matching questions found. Try a different search term.

CONVERSATION GUIDE

How to sell {{product.name}}.

Pick who you are talking to and how far the client wants to go. You get the message that lands with that buyer and the matching entry point on the offer ladder. Pricing is not on this page: scope and price come from your commercial lead.

BUYER — Which persona?

Who are you talking to?

ENTRY POINT — Which offer?

How far do they want to go?

Select options on both sides to see your recommendation

Persona

Entry point

PERSONA MESSAGING

What to say

Where it applies

Typical client questions

What we need from the client

Best-fit customers

Indicative timeline

OFFER

Scope and terms

What drives scope

What is included

PLAYBOOK

Pitch, position, handle, qualify.

Drawn from the product brief. Competitor statements are factual summaries of what each alternative does, not claims about their quality: have Legal review before any of this goes into a client document.

ELEVATOR PITCH

Say it in 30 seconds.

30-second version

"For CFOs and chief AI officers who are spending real money on AI but cannot say what it is buying, {{product.name}} is a {{product.category}} managed service that connects to your AI providers' billing, shows a cost baseline with an honest score for how much of the spend it can see, and ties that spend to business outcomes you have signed off. Unlike a gateway or a cloud cost tool, which show spend by request or by account, we link cost to outcomes and keep the working visible behind every number. The first step is a four-week Diagnostic. Approved references are shared on request."

Tips

Lead with what the client cannot answer today, not the technology.
If you have no cleared proof point, say nothing rather than borrow one.
End with the Diagnostic as the next step.

Two-minute version: structure

Context: 98% of FinOps practitioners now manage AI spend directly, up from 31% in 2024 (FinOps Foundation, State of FinOps 2026). Token prices stopped falling around November 2025, so waiting for costs to shrink is no longer a plan.
Problem: Blind Overspend. The invoice is visible, but not cost per use case, cost per outcome, or whether the checking humans do is worth what it costs. Agents add spend that no control catches.
Solution: read-only connections to every AI provider, a baseline with a completeness score, outcomes signed off by the client, and agent sessions authorised before they start.
Proof: show, do not claim. The demo tenants make the point in minutes. Approved client references are shared on request once cleared; until then, use the demo tenants.
Next step: a Diagnostic scoping call.

POSITIONING

Where {{product.name}} fits.

We are the right choice when

The client uses more than one AI provider, or runs AI inside shared cloud subscriptions where the bill cannot be tied to a use case.
Finance has asked for cost per use case or per outcome and nobody can answer.
The client runs, or is about to run, autonomous agents in production.
There are not enough FinOps-for-AI specialists to staff this by hand: that is true market-wide.

Named alternatives: what they do and where we sit

LiteLLM, Cloudflare AI Gateway, Portkey, Martian: request-level routing, deciding which model handles a call at what latency and cost. We extend this layer (the gateway substrate is LiteLLM). We do not compete with it.
Azure AI Foundry, Amazon Bedrock Guardrails: policy enforcement, checking whether a request complies with a rule set. We sit alongside. Compliance is not value.
AWS autonomous FinOps agent: finds and executes savings inside one cloud. We are provider-neutral and sit above the routing layer.
Revenium: frames itself as an "AI economic control system". Our difference is the calibration methodology and the link from cost to signed outcome value, not the dashboard.
Internal FinOps team plus spreadsheets: manual attribution. There is not enough FinOps-for-AI talent in the market to do this by hand at demand.
Do nothing: unit costs are no longer falling, so the gap between spend and understanding widens.

Rule of thumb: any screen a gateway vendor could ship as a free feature (token spend by key) is not our differentiator. Outcome linkage, oversight calibration and agent containment are.

Not a fit, or not yet

The client wants a model router or gateway only: that is the substrate we extend, not what we sell.
The client wants prompt or completion content stored or analysed: not in the product today. Route to the Data Privacy Officer and the offer owner.
The client asks whether the platform meets a specific regulation: route to Legal and Compliance. Do not answer from the product.
The client expects Verified Yield immediately: it needs a captured baseline and belongs to Tier 2 onward.
The client cannot grant read-only billing access to any provider: there is nothing to baseline.

Saying "not yet" early protects credibility and pipeline quality.

OFFER LADDER

Six ways in, no prices on this page.

Diagnostic

Four weeks, fixed scope.
Cost baseline with completeness score, unit economics, top five optimisation opportunities, governance framework.
No long-run monitoring. The usual first step.

Tier 1: Govern

Cost baseline and governance artefacts.
Checkpoint gates G1 to G3.
Identifies opportunities; does not execute them.

Tier 2: Attribute

Adds the CPVO attribution engine and signed outcome definitions.
Gates G1 to G5.
Verified Yield, once a pre-AI baseline is captured.

Tier 3: Optimise

Adds continuous routing and rightsizing recommendations.
Gates G1 to G7.
Every change needs a recorded human approval before it is actioned.

Control Plane (roadmap)

Self-service tier: coverage of AI embedded in SaaS products, prompt coach, risk classifier.
Not generally available: the connector surface exists but no SaaS adapter is live. Describe it as roadmap only, and confirm current status with the product team before mentioning it to a client.

Agentic add-on

For clients running agents, on top of any tier.
Pre-launch authorisation (Gate A), in-flight monitoring, session audit.
Priced separately.

⚠

Scope and price are agreed with the Commercial Director. The figures in the product brief are indicative and not validated, so do not quote, hint at or compare them. Each offer is scoped per engagement.

OBJECTION HANDLING

Hear it, acknowledge it, answer it.

"We already have a gateway / cloud cost tool."

Acknowledge: good, that is the layer we build on.
Answer: those tools show spend by request or by account. We tie it to a use case and a signed outcome, and say how much of the spend we can see.
Ask: "Can you tell me today what one resolved ticket costs you?"

"How do we know it is secure and compliant?"

Answer: the connections are read-only (billing and numeric usage, no ability to invoke models), credentials are encrypted at rest and never sent to the browser, and prompt and completion content is not stored.
A Data Privacy Officer assessment is required before any real client is onboarded. Do not claim certifications or regulatory fit yourself. Ask the product team for the approved security summary when one is available.

"It is too expensive / not budgeted."

Answer: start with the four-week fixed-scope Diagnostic: a small first step that produces a baseline and a ranked list of opportunities.
Do not quote prices from this page. Use your commercial lead.

"We do not have the people or time."

Answer: Capgemini runs the checkpoint operating model. The client provides read-only access, a named signatory for outcome definitions, and outcome counts from their own systems.
L1 consultants need no FinOps-for-AI specialism: the platform prepares the analysis, a person chooses from bounded options.

"Prove it works for us."

Answer: the Diagnostic builds the baseline from their own billing, and the completeness score shows what it could not see.
Use no client reference unless it has been cleared.

"Show me the formula."

Answer: you get the inputs behind every figure: the cost rows, outcome counts and the version in force. The calculation engine and its parameters stay with Capgemini.
Do not describe or estimate the parameters. Disclosure of protected IP needs Chief AI Officer sign-off.

"What savings will we get?"

Answer: the Diagnostic lists addressable amounts per finding, marked as inferred until validated with the client. They are not committed savings.

"Will it lock us in to a cloud or model vendor?"

Answer: no. It reads from Azure, AWS, Google Cloud, OpenAI and Anthropic through one normalised schema, and works with the client's existing gateway.
Hosting region and data residency: the application runs on Microsoft Azure Container Apps in North Europe. Tenants carry a data region property, and a dedicated deployment can be configured where a client needs a specific region. Confirm the wording with the Data Privacy Officer before quoting it.

"Can it stop a runaway agent?"

Answer: the design is a budget authorised before launch, a monitor that pauses then stops a session, and a full audit trail. Be exact about status: see the Demo tab on what can be shown today.
Do not promise live enforcement at the gateway until the product owner confirms it. Live pause and stop of a running agent at the gateway is not confirmed today.

QUALIFICATION

Is this a real opportunity?

Need

Which AI providers do you use, and who owns each bill?
Can finance tell you cost per use case today?
Are agents running, or planned, in production?
What happens at the next board review if the question comes up?

Authority & budget

Is the CFO involved, and who is the delivery sponsor (CAIO or CTO)?
Is there budget, or a route to it, for a four-week first step?
Who would sign what one outcome is worth?

Readiness & timing

Can you grant read-only billing access to each provider, and how long does that take?
Does the client system of record count the outcomes (tickets, claims, policies)?
Are security, procurement and data protection steps known?

⚠

Qualified? Tell the {{product.name}} team at {{product.contactEmail}} so a scoping call can be arranged. Log it in the deal tracker with an honest stage and probability. Keep client-confidential details out of this site.

Pitch deck

The current introduction deck: the story, the model and proof points. Use it to brief, or leave it behind (check the client may receive it first).

Open the deck

Team tool

Deal tracker

The team pipeline: log new conversations, tag hot deals and services, and see where the weighted value sits.

Open the deal tracker →

Show it

Demo video & scripts

The recorded walkthrough and live-demo scripts are on the Demo path.

Go to Demo {{product.name}} →

Demo video

{{product.name}} product demo

A short, self-running walkthrough: the fastest way to make the product concrete in a conversation.

Watch the demo

Pitch deck

The current introduction deck, to talk over or leave behind.

Open the deck

Team tool

Demo hub

Find, book and track the shared demo environments, and see who has which one and until when.

Open the demo hub →

⚠

Where the target moment is not met yet. (1) A credential can be connected and verified in the meeting, but first billing data arrives on the next scheduled sync, every four hours, unless a platform admin triggers a run; only platform admins can add connections. (2) The agent sessions on screen are seeded. There is no controllable looping test agent in the product, and live pause and stop of a real running agent at the gateway is not confirmed. Show the control and the record it leaves; do not claim you are stopping a live agent. Update this note when either gap closes.

Choose a format

Three formats, built from the same four scenarios.

5 min

Lightning demo

Scenario 1 only: the cost baseline with its coverage badge. For a CFO in a corridor or a short intro.

Environment: a demo tenant, in client mode

20 min

Standard walkthrough

Scenarios 1 to 3: baseline, cost per verified outcome, agent controls. With questions.

Environment: Northwind, Meridian, then Vantage

60 min

Working session

All four scenarios, including a connection set-up with the client's own platform team and a look at the checkpoint console.

Environment: demo tenants plus a prepared connection

The demo tenants

Three fictional tenants are seeded in the app and carry a yellow DEMO badge in the top bar. They are not real clients.

Northwind SaaS demo-northwind-saas

A mid-market SaaS company. Uniformly healthy: every workload computes. Instrumented only three weeks ago, so it is the one that shows the split between traffic seen at the gateway and traffic that was not. Two pilots are measured by cost per run. Use it for Scenario 1.

Meridian Insurance demo-meridian-insurance

A regulated insurer. Deliberately shows "not attributable" (an unsigned outcome definition), a coverage caveat and a low-confidence figure, and is heavy on checkpoints and escalations. Two applications sit off the gateway, so two of its three billing scopes carry a real gap. Use it for Scenario 2 and the checkpoint console.

Vantage Agentic demo-vantage-agentic

An agent-heavy technology company with a spread of healthy, paused and stopped sessions and a tripped circuit breaker. A mature gateway: five scopes, all with small gaps. Use it for Scenario 3.

Sign-in and mode before you start

Sign in with your Capgemini account at the AI Yield app. You need an account provisioned by a platform admin (Build tab). Use the mode pill in the top bar to switch to client mode before sharing your screen. Spend on the demo tenants is deliberately in the low thousands a month: say figures are illustrative.

⚠

Synthetic data on the demo tenants only. Never load real client data, credentials or personal data into a demo tenant. The demo tenants share a production database with the real internal tenant, so work only inside tenants carrying the DEMO badge. Make clear that demo figures are illustrative. If you are unsure whether data is permitted, ask the Data Privacy Officer first.

Prepare: know your audience

Find out who attends. For a CFO, build the demo around Scenarios 1 and 2. For a CAIO or CTO, lead with Scenario 3 and the checkpoint console. Ask which providers they use so you can say which are covered and what each needs. Write down the three things you want them to remember.

Rule of thumb. If you cannot state the client's problem in one sentence, go back to discovery.

Set up the day before

Sign in, check your role, and open each demo tenant. Switch to client mode and walk the whole script once. Confirm the pages load and the DEMO badge shows. If you plan to connect a client credential live, agree with a platform admin that they will be on hand, and read the in-app "What do I need?" guide for that provider with the client's platform team beforehand.

Fallback. No recording is published yet; ask the product team whether one is available. If the app is slow, switch to screenshots of the same pages and say so. Do not improvise numbers.

Open (first 2 minutes)

Restate the client's problem in their words, say what you will show and how long it takes, and invite interruptions. Do not start with a tour of the navigation.

Script

"You told us you cannot say what your AI spend is buying. In the next [N] minutes I'll show how {{product.name}} gets from the provider bill to a cost per outcome, how it tells you what it cannot see, and what it does about agents. The data is a fictional company. Stop me any time."

Core storyline: four scenarios

Same rhythm each time: situation, what you show, what to say, and the "so what" for the client.

Scenario 1: a real cost baseline with an honest score about 6 min, Northwind

Show: open Govern. The header figure is the spend for the period, and the coverage badge sits next to it, not on another tab. Walk the breakdowns: by provider, by workload, by model family. Point at the unattributed spend, which is ranked in place and never spread across workloads. Then Optimise, Gaps: the billing scopes where invoiced spend exceeds what the instrumented applications account for. Say: "This is what we can see, and this is what we cannot. The gap is the next action." So what: a CFO's credibility with the board starts with an honest coverage figure. Live connection variant: open Connections, add a read-only credential for a provider the client uses (platform admin only). Show it verify and display the granted scope: read-only billing and numeric usage. Then explain that data arrives on the next sync, and switch to a pre-synced tenant. For AWS the credential needs only two actions, cost query and metric read, and cannot see workload content. If a platform admin triggers a sync during the meeting, show the baseline appearing; do not promise it.

Scenario 2: cost per verified outcome, and when it refuses about 6 min, Meridian

Show: Attribute. Open a workload that computes and use the provenance link to show the cost rows, outcome count and the version behind the figure. Then Customer Correspondence: "not attributable", because the outcome definition is not signed. Then Underwriting Risk Summary (coverage caveat) and Regulatory Reporting (low confidence, too few verified outcomes, with the count shown). Show Definitions, where a definition needs a client and a Capgemini signature. If a workload with a baseline on file is available, open Yield. Mention the Call Centre Sentiment Monitor, which roughly breaks even: that is what Optimise points at. Say: "We would rather show you a blocked state than a number that fails the first challenge." So what: every figure survives a CFO's questions because the inputs are one click away. Do not describe how cost or value is weighted, and do not call anything Verified Yield unless a baseline exists for that workload.

Scenario 3: what stops an agent running away about 6 min, Vantage

Show: Agentic. Sessions are sorted by risk, so the ones needing attention are at the top, each with a budget bar, steps and the signal that triggered it. Open a paused session: the decision the platform recorded when it authorised it, the signals, and the actions taken with timestamps. Show Policy (budget, step and external-action limits per workload), History (filterable, exportable audit) and Economics (the share of agent spend consumed by sessions that never finished). Show the tripped circuit breaker. Say: "A session cannot start without an approved budget. If it loops, it is paused, and if it keeps going, stopped. An overspend stop has no override at first-line level, and the record shows what the platform knew at each step." So what: this is the audit artefact a regulated client asks for. Be exact: these sessions are seeded. Say "here is the control and the record it leaves", not "watch it stop a live agent". The screen labels states with colour names: say paused and stopped.

Scenario 4 (CTO audience): how it runs without specialists about 8 min, Meridian

Show: sign in as a user with a checkpoint role and open the console (internal screens, so do not share them in client mode with client data you have not cleared). Open one checkpoint: the question, the finding, the confidence signal and its drivers, the bounded options, the timer and the escalate button. Show the gate sequence G1 to G7 and an escalation with its clock. Say: "A generalist chooses among options in about twenty minutes. Experts see only exceptions." So what: the delivery model scales without a specialist on every engagement.

Handle questions and "can it do X?"

Answer what you know. For roadmap items (Control Plane, SaaS-embedded AI coverage, live gateway enforcement), say so and promise a follow-up. Never demonstrate or promise an unreleased feature as available. Route security, privacy and regulation questions to the Data Privacy Officer and Legal and Compliance. Park deep technical questions for a session with an architect.

Close and follow up

Recap the three things you wanted them to remember and agree one next step: normally a Diagnostic scoping call and a list of the providers and read-only access involved. Send a short follow-up within 24 hours and tell the product team at {{product.contactEmail}}. Update the deal tracker.

Clean up

Demo tenants should look the same for the next presenter: do not add, sign or delete anything in them. If something was changed by accident, a platform admin can reseed a single tenant (make seed-demo with ONLY=<slug>), which deletes and recreates it. Remove any test connection you created. Note what worked so the next presenter benefits.

Open the demo hub →

Step 1

Go to the AI Yield app and sign in with your Capgemini account through Microsoft Entra ID.

Step 2

Get a role

A first sign-in shows "Access not yet provisioned" until a platform admin adds your account and role. Ask your capability owner, or write to {{product.contactEmail}}. The approver is a platform admin on the product team; allow a working day or two, and confirm the turnaround with them.

Step 3

Open a demo tenant

Pick one of the three DEMO tenants from the engagement list. Do not use any other tenant in front of a client without checking with the engagement team.

Roles what each can reach

platform_admin: everything, including user admin, calibration, all engagements and adding provider connections.
capability_owner (L3): engagements, checkpoints, escalations and calibration.
expert (L2): checkpoints and escalations.
consultant (L1): checkpoints.
commercial: the home page only.
client_admin, client_operator, client_viewer: the client's own workspace, with administrator, operational or read-only rights. They never see calibration or thresholds.

Who to contact by question

Product, access and demos: {{product.contactEmail}}. The product owner is the {{product.name}} product lead; ask for the team channel when you are given access.
Personal data, privacy and regulated data: the Data Privacy Officer.
Legal, compliance, regulation: Legal and Compliance.
Pricing and scope: your commercial lead and the Commercial Director.
Disclosing how the calculation works: the Chief AI Officer, with a dated and witnessed disclosure document.
Deployment and infrastructure problems: whoever holds platform admin on the hosting resource group.

Links for builders

The AI Yield app
Technical enablement portal
Source repository (README, runbook, product brief pack and design decision records). Repository access is separate: ask the product team.

Now · Live

Release {{product.version}}

What exists in the app

Live capabilities

Read-only connections with verify-on-save to Azure, AWS Bedrock, Google Vertex AI, OpenAI and Anthropic, synced every four hours
Govern: cost baseline, hidden cost drivers, coverage score with named gaps, exports
Attribute: workloads, signed outcome definitions, cost per verified outcome states, Verified Yield with a baseline, AYCI
Agentic: Gate A authorisation, policy, session monitor, history and economics views
Optimise: gaps, opportunities and recommendations with authorisation
Checkpoint console G1 to G7, escalations, calibration, dual-mode, three demo tenants

Next · timing to be confirmed · preliminary scope

Close the known gaps

What we are building now

New & evolving capabilities

Live pause and stop of a running agent at the gateway, and a controllable test agent for demos (not yet built; timing to be confirmed)
Sync on demand, so a baseline can appear in the meeting after a credential is connected (planned; timing to be confirmed)
Billed (not estimated) Vertex cost, which needs a Google Cloud billing export
Real Bedrock spend seen end to end on a client account

Platform themes

Instrumentation for checkpoint duration, escalation rate and platform run cost, which the commercial validation depends on

Later · On the horizon

Direction of travel

Where we want to go

Candidates

Control Plane: self-service onboarding, coverage of AI embedded in SaaS products, prompt coach, risk classifier
Native AI token fields as the FOCUS 1.5 standard lands (confirmed for December 2026)
Cross-client benchmarks, only above a floor of five tenants and with Legal sign-off

Roadmap items are indicative and may change. "Next" scope is preliminary; "Later" is direction, not a committed delivery date. Do not present unreleased items to clients as available. Release notes: enablement portal.

01 · Architecture

Architecture overview

System context, major components, request lifecycle and data tiers.

02 · Components

Components & capabilities

A catalogue of the building blocks and what each one can do.

03 · Features

Feature reference

Every feature: what it does, how it works, how it is configured and how it links to others.

04 · Operations

Operations

Run-time operation: monitoring, support, administration and troubleshooting.