Skip to content
AI Governance

The line item that grew up overnight.

Twelve months ago AI spend was a footnote. Now it's the fastest-growing line on your P&L, with the loosest controls. Flowstate gives you the tools to take it back — set the rules, watch what's happening, and stop the leaks before they leak.

Policy events · last 60 min
Live
Approved
Sarah K. · Payments v2
£14.50$18.30
Substituted
chatbot‑prod · Opus → Sonnet
£0.74$0.93
Blocked
m.harris · marcusphd/thesis
£42.00$53.00
Approved
Video Generation · Service request
£0.21$0.27
Soft-cap
d.adams · £203 / £200$256 / $252 personal cap
+£3.00$3.78
Voice of the customer

The same conversation, in every boardroom we sit in.

Lightly anonymised, from a year of CFO and CTO conversations.

I have no visibility into where AI investment is actually going — or whether it’s working.
CFO · public consumer brand
I would pay anything to know AI costs per project, and what an investment would unlock.
CTO · fintech
Controlling our AI spend is one of the greatest challenges we have to solve.
SVP Engineering · professional networking
We’re wasting millions by not optimising where AI spend goes and how we deploy it.
CIO · global enterprise software
Worked example

One project. Four questions. Four answers.

Take the Video Generation feature your team shipped in March. Twelve weeks of build. Three engineers. £12,150$15,300 of OpenAI tokens to prove the architecture. Here’s how Flowstate answers each of the questions a CFO would ask about it.

01 — Attribute

"Whose tokens were these?"

The OpenAI invoice says £12,150$15,300. The provider can tell you it hit one API key. Beyond that, it’s guesswork — until each request is tied back to a person, a service and a project.

Video Generation · attributed spend 12 weeks
SK
Sarah K. · dev
£4,820$6,070
PM
Priya M. · dev
£3,240$4,080
JT
James T. · dev
£1,690$2,130
vg-eval-pipeline · service
£2,400$3,020
Total attributed £12,150$15,300 100%
02 — Classify

"Is any of this capitalisable?"

Video Generation was in development phase for the whole twelve weeks. That means the £12,150$15,300 isn’t just OpEx — most of it is capitalisable, and a chunk of that qualifies for R&D tax relief. Flowstate splits it automatically because we already know the project stage.

Accounting treatment Auto-classified · lifecycle: development
CapEx
R&D
OpEx
CapEx
£8,630$10,860
Build phase · FRS 102ASC 350-40
R&D eligible
£2,180$2,750
Tech‑uncertainty work
OpEx
£1,340$1,690
Eval · pre-launch ops
R&D narrative pre-authored. "Novel evaluation harness for diffusion-based video generation; iterative resolution of token-budget under deterministic seed control..." — ready to review.
03 — Forecast

"What’s it going to cost when this scales?"

Now Video Generation is live. Adoption is climbing. There’s no seat-based ceiling like Copilot used to give you — production AI scales with traffic. Your CFO needs a number for the next board meeting.

Forecast · next quarter
£94,200$118,500
Range £81k – £109k$102k – $137k at 80% confidence
Trend
+47% QoQ
Actuals
Forecast
Mar
Sep
Scenario: swap two contractors for an autonomous review agent next quarter. Net P&L delta: £42,000$53,000.
04 — Control

"How do we stop the leaks before they leak?"

Visibility is the start. Real control means policy you can write down, enforcement that fires when it’s breached, and a graduated response — a quiet nudge first, a manager surfacing later, an evidence pack only if it escalates.

Video Generation · live policy
v3 · owned by chris.eng
Approved models: GPT-4o-mini, Sonnet, Gemini Flash
Per-customer-action cap: £0.40$0.50
Anomaly detection: traffic shape, request size, model drift
Graduated response
Tier 1
Quiet nudge
Slack DM
Tier 2
Soft cap
Off-policy spend rate-limited
Tier 3
Manager surfacing
Conversation, not report card
Tier 4
Evidence pack
Defensible escalation
The toolkit

What you actually get

Concrete controls you can wire up. Pick what fits your culture; layer the rest as you grow into it.

Set the rules

Guidelines

Soft, educational nudges. Recommended models, suggested providers, healthy budget norms. The polite first conversation, not a block.

Policies

Hard rules with owners and version history. Approved models, per-team caps, allowed providers per service, capitalisation rules per project stage. Versioned, audited, exportable.

Watch what’s happening

Activity log

Every AI session, every API call, every policy decision — recorded, searchable, filterable, exportable. The audit trail you’d want to hand to a board or a tax investigator.

Alerting

Real-time alerts to Slack, email or your incident channel. Policy violations, anomalous usage, budget drift, unit-cost ceilings being breached.

Stop the leaks

Budget cut-off

Hard ceilings per person, team, project or service. When spend hits the cap, requests stop. Soft caps too.

Auto banning

Repeat policy violations trigger temporary suspension. Manager surfacing. Manual unblock.

Key revocation

One-click rotation or kill. Compromised key revoked across the whole stack in seconds.

Anomaly response

Service identity verification — if a chatbot key starts being used for something else, the request is rejected.

One model, two faces of AI spend

Developer AI and production AI look different on the surface. The economic questions are the same, so the model should be too.

Developer AI
Interactive · per-seat + overage · tied to people

Claude Code, Cursor, Copilot, Windsurf. The questions are about projects, productivity, and which engineer is suddenly spending £4k$5k a month on Claude.

This month, Platform team £18,420$23,210
Sarah K. £3,840$4,840
Marcus H. £2,910$3,670
Priya M. £2,610$3,290
Marcus reaching for Opus on routine work · £420$530 potential saving / mo
Production AI
Autonomous · per-token · tied to services

Chatbots, document extractors, video pipelines. The questions are about contribution margin per customer action, and why your OpenAI bill doubled in May.

This month, AI services £62,840$79,180
customer-chatbot
£0.04$0.05 / action £28,420$35,810
video-generation
£0.31$0.39 / action £21,180$26,690
doc-extractor
£0.08$0.10 / action £9,140$11,520
summariser-v1
− margin £4,100$5,170
summariser-v1 above unit-cost ceiling for 9 days · service owner notified

How the data lands

Three ways to bring AI activity into Flowstate. Pick the ones that fit how your org works — you don’t need all three to start.

Provider integrations

Direct billing & usage from each AI provider. Reconciled nightly. Nothing to install.

Anthropic OpenAI Gemini Cursor Windsurf

Open-source telemetry

Lightweight CLI you push via Homebrew or your MDM. Detects whichever AI tools your engineers run. Sends features back — never source code.

Jamf Intune Kandji

Stop spend in flight

For services and teams where policy needs to fire mid-request — rate limits, model substitution, mid-session caps, hard cut-offs. Optional.

Flowstate ledger
reconciled nightly
One source of truth across people, projects, services, finance and AI spend.
01
Attribute — tokens to person/service/project
02
Classify — CapEx, OpEx, R&D-eligible
03
Forecast — quarter-ahead, by team and project
04
Control — policies, alerting, budget cut-offs

Your contracts, your keys, your data

An insight and policy layer — not a reseller, not a billing middleman.

No vendor lock-in

Your provider contracts stay yours. Change vendors whenever — Flowstate picks up the new one through the same integrations.

No billing intermediation

Your Anthropic invoice still comes from Anthropic. Your OpenAI invoice still comes from OpenAI. We don’t touch the money.

Prompts stay private

We hold metrics and attribution. Prompt and response bodies are processed in memory and discarded — they never persist on our infrastructure.

Stop flying blind. Then stop the leaks.

Book a demo and see exactly where your AI budget is going — across developer AI and production AI — and what you can actually do about it.