Approach

AI runs the operation. Named people make the decisions.

Agents draft, check, localise and monitor; automatic tests hold their work to your standard. A named person approves each step that goes live, and every action is logged for you to read.

Try the demo Compare the packages
Stages
Discover → Improve
Approval gates
One per stage
Audit log
Every action

A sample operations console: agents have drafted twelve campaign variants, eleven passed the brand checks and one unsupported claim was held. The work waits for the brand lead to approve or send back; each step is logged with a time.

How we use AI

Five controls between an agent and your brand.

Agents bring the speed; the other four controls make their work fit to approve. All five apply to every project we run.

ControlWhat it doesIn the logWho owns it
01Agents Do the repeatable work: research synthesis, drafts, variants, code changes, checks and monitoring. agent.localise · 4 markets · 38 assets Built and run by us
02Evaluations Score every output against a written standard, so work that fails never reaches an approver. eval.tone · 0.92 ≥ 0.85 · pass Thresholds set with you
03Automatic checks Hard limits an agent cannot cross: unsupported claims, personal data, off-brand language and prompt injection. guard.pii · 2 fields redacted Rules owned by you
04Approval gates A named person approves each step that goes live, spends money or speaks for you; the agent waits. gate.legal · approved · A. Rao Your approvers
05Audit logs Every prompt, model, output, score and decision is written down with who, what and when. log · 1,284 entries · exportable Readable by you, always
A woman at a laptop taking a printed document from a colleague standing beside her desk.
Named approverAn agent drafts; a named person reads it and signs it off before it goes live.

People decide what to make, for whom, what it may claim and whether it goes live.

How it works

How a job moves from request to approval and release.

Choose a sample job and watch it move. At the human gate, approve it or send it back with a note for the agent to redraft.

  1. 01Signal

    Drop-off spike at payment

    Analytics alert · step 3 · mobile

    alert · p75 INP 310 ms

  2. 02Agent

    Coding agent

    Proposes a fix in a branch with tests

    PR · 4 files · 2 tests

  3. 03Evals

    Test + performance evals

    Unit, e2e, Lighthouse, INP budget ≤ 200 ms

    INP 160 ms · all green

  4. 04Guardrails

    Security guardrail

    Dependency scan, secrets scan, OWASP checks

    0 critical · 0 secrets

  5. 05Human gate

    Tech lead reviews

    Human code review, then a staged rollout

    Approver · tech lead

  6. 06Ship + learn

    Rolled out at 10% → 100%

    Automatic rollback if error rate rises

    drop-off watched 14 days

Audit log append-only

  1. signal.alert · checkout step 3 · INP 310 ms
  2. agent.code · branch fix/pay-input
  3. eval.tests · 212 pass · INP 160 ms
  4. guard.security · scan clean
  5. gate.review · approved · 1 comment resolved
  6. ship.rollout · 10% canary
  7. ship.rollout · 100% · rollback armed

Sample runs for a fictional “Your company”. Timings are illustrative.

Delivery stages

Six delivery stages, each signed off by a named approver.

Each stage sets out what agents do, what you keep and the written exit criteria your approver checks before signing off.

  1. 01

    Discover

    Typically 1–3 weeks

    What is true today, and what is it costing you?

    Agents do

    Turn interviews, analytics, reviews and competitor data into themes, each with its sources.

    You keep

    • Evidence map with sources
    • Stakeholder interview notes
    • Opportunity shortlist

    Gate · Problem framedExit: Agreed problem statement and success measures

    ApproverYour executive sponsor

  2. 02

    Define

    Typically 1–2 weeks

    What will we build, and how will we know it worked?

    Agents do

    Draft requirement options, estimate effort ranges, flag risks and dependencies from the evidence.

    You keep

    • Scope and acceptance criteria
    • Measurement plan
    • Risk and dependency register

    Gate · Scope signedExit: Scope, budget model and measures signed

    ApproverSponsor + product owner

  3. 03

    Design

    Typically 2–6 weeks

    Does it work for the people who will use it?

    Agents do

    Draft variants within your brand guidelines and run accessibility and content checks on every screen.

    You keep

    • Prototype and design system tokens
    • Accessibility review (WCAG 2.2 AA)
    • Usability test findings

    Gate · Design approvedExit: Tested prototype with no open WCAG AA issues

    ApproverProduct owner + brand lead

  4. 04

    Build

    Typically 4–16 weeks

    Is it correct, fast and secure?

    Agents do

    Propose code with tests, then run evaluations, performance checks and security scans on every change.

    You keep

    • Working software in staging
    • Test and evaluation reports
    • Threat model and security scan

    Gate · Release approvedExit: All checks passed, code reviewed by a person

    ApproverTech lead + your IT/security

  5. 05

    Run

    Typically Ongoing

    Is it staying healthy after launch?

    Agents do

    Monitor uptime, Core Web Vitals, cost, off-brand output and stale content; open tickets with evidence.

    You keep

    • Runbook and on-call rota
    • Live dashboards
    • Incident and change log

    Gate · Service acceptedExit: Service levels met for an agreed period

    ApproverYour operations owner

  6. 06

    Improve

    Typically Quarterly

    What should change next, and why?

    Agents do

    Rank experiments by expected value, draft variants and feed results back into the backlog.

    You keep

    • Experiment log
    • Quarterly outcome review
    • Updated roadmap

    Gate · Roadmap resetExit: Results reviewed and next experiments agreed

    ApproverSponsor, quarterly

Quality bars

Four quality bars every release must meet.

Written into acceptance criteria and checked automatically on every change. These are frameworks we build to, not certifications we hold.

Performance · Core Web Vitals

“Good” at the 75th percentile of visits

  • LCP≤ 2.5 sLargest Contentful Paint · loading
  • INP≤ 200 msInteraction to Next Paint · responsiveness
  • CLS≤ 0.1Cumulative Layout Shift · visual stability

Core Web Vitals — Loading, interactivity and visual stability

Accessibility

WCAG 2.2 AA, tested by tools and by people

  • 4.5:1 text contrast, 3:1 for large text and UI
  • 24 × 24 CSS px minimum targets (2.5.8)
  • Focus never hidden behind sticky UI (2.4.11)
  • Keyboard and screen-reader passes before release

WCAG 2.2 AA — Web Content Accessibility Guidelines

Security and privacy by design

Threat-modelled before it is built

  • ASVS controls chosen per risk, verified in review
  • LLM01 prompt injection tested on every AI feature
  • Data minimised, purpose-bound, consent recorded (DPDP, GDPR)
  • Secrets scanned on every commit

OWASP ASVS — Application Security Verification StandardOWASP Top 10 for LLM Applications — Security risks in generative AI applicationsDPDP Act 2023 — Digital Personal Data Protection Act, 2023

Carbon · Software Carbon Intensity

Measured per unit of use, then reduced

SCI = ((E × I) + M) per R

E
Energy used by the software, kWh
I
Carbon intensity of that energy, gCO₂e/kWh
M
Embodied emissions of the hardware share
R
The functional unit: a visit, a call, a user

SCI · ISO/IEC 21031:2024 — Software Carbon Intensity

How we work with you

Weekly reviews, one backlog and a decision log.

You work from the same board and scores as our team, and every decision is written down with its reason.

  1. MonPlan

    Backlog triage

    Agents pre-rank new items by value and effort; the product owner sets the order.

  2. TueBuild

    Working sessions

    Joint sessions on the hardest problems, while agents test each change as it comes in.

  3. WedCheck

    Risk and quality

    Test scores, held items and budgets reviewed; anything failing gets a named owner.

  4. ThuBuild

    Demo prep

    Only working software or finished work goes into the demo.

  5. FriDecide

    Weekly review

    Demo, decisions recorded and next week agreed, in thirty to forty-five minutes.

What you always have access to

  • Shared backlog. One list, ranked by your product owner, visible to everyone on both sides.
  • Decision log. What was decided, why, by whom and what it replaced.
  • Quality board. Test scores, Core Web Vitals, open risks and cost, updated live.
  • Audit log export. Every agent action, on request, in a format your auditors can read.
Seen from above, four colleagues around a table covered in printed charts, one pointing at a pie chart.
FridayThe weekly review: the same charts on both sides of the table, decisions written down before anyone leaves.

Decision log · Your company · sample

IDDecisionBecauseOwnerState
D-041Launch 3 markets first, 1 laterLegal review time in market 4SponsorApproved
D-042Raise tone eval threshold to 0.85Two drafts passed that read off-brandBrand leadApproved
D-043Keep human review on all refund answersLow volume, high riskSupport leadApproved
D-044Move search to a hosted vector indexCost vs latency trade-offTech leadOpen

Responsible AI

Six commitments we write into the contract.

Each commitment is enforced on one lane of the board above, in the same order, and written as a contract clause you can hold us to.

Frameworks we build to

ISO/IEC 42001:2023 — Artificial intelligence management systemsNIST AI RMF 1.0 — AI Risk Management FrameworkEU AI Act — Artificial Intelligence Act (EU) 2024/1689OWASP Top 10 for LLM Applications — Security risks in generative AI applicationsMITRE ATLAS — Adversarial Threat Landscape for AI Systems

  1. Your data stays yours

    Client data is never used to train shared models. Retention is set per project and written down.

    Enforced on the board lane: Signal signal.* · your data, your retention

  2. Disclosed

    You always know which work an agent touched, which model it used and which prompt version.

    Enforced on the board lane: Agent agent.localise · model + prompt v4 logged

  3. Checked for accuracy and bias

    Tests for accuracy, bias and harmful output run before launch and on a schedule afterwards.

    Enforced on the board lane: Evals eval.grounding · 1.00 · pass

  4. Attack-tested

    Tested for prompt injection, data leakage and jailbreaks, mapped to the OWASP Top 10 for LLM Applications and MITRE ATLAS.

    Enforced on the board lane: Guardrails guard.llm01 · input clean

  5. Human-accountable

    Every agent has a named human owner. No agent approves its own work.

    Enforced on the board lane: Human gate gate.market · approver · market lead

  6. Reversible

    Every automated action can be paused, rolled back or switched to manual by your team.

    Enforced on the board lane: Ship + learn ship.rollout · rollback armed

Close-up of a hand signing a printed form with a fountain pen on a light wooden desk.
Human-accountableNo agent signs for itself. Each one works under the signature of a named owner who answers for what it does.

Engagement packages

Six packages, all with the same controls.

Packages differ in which delivery stages they cover, how much is fixed up front and how long we stay.

  1. 01

    Sprint

    A short, fixed-scope engagement that answers one defined question.

    Typical length
    1–3 weeks
    Pricing model
    Fixed fee
    Best for
    Discovery, a diagnostic, a prototype or a decision you need to make soon

    Stages covered

    1. Discover — covered
    2. Define — covered
    3. Design — not covered
    4. Build — not covered
    5. Run — not covered
    6. Improve — not covered

    Ends at the “Problem framed” or “Scope signed” gate, with a decision in hand.

    Enquire about a sprint engagement
  2. 02

    Project

    A defined scope, delivered for a fixed price.

    Typical length
    4–12 weeks
    Pricing model
    Fixed price
    Best for
    Work you can describe up front: an identity, a platform or a set of tools

    Stages covered

    1. Discover — not covered
    2. Define — covered
    3. Design — covered
    4. Build — covered
    5. Run — not covered
    6. Improve — not covered

    Scope signed up front; ends when the release gate is approved.

    Enquire about a project engagement
  3. 03

    Milestone

    A larger build in phases you approve and pay for one at a time.

    Typical length
    3–9 months
    Pricing model
    Fixed price per milestone
    Best for
    Programmes too large for one contract, where you want control at each step

    Stages covered

    1. Discover — not covered
    2. Define — covered
    3. Design — covered
    4. Build — covered
    5. Run — covered
    6. Improve — not covered

    You approve, and pay for, one gate at a time; stop after any of them.

    Enquire about a milestone engagement
  4. 04

    Retainer

    Reserved monthly capacity to run, improve and extend what we built.

    Typical length
    Ongoing · 6-month minimum
    Pricing model
    Monthly fee
    Best for
    Live brands and products that need a steady team without hiring one

    Stages covered

    1. Discover — not covered
    2. Define — not covered
    3. Design — not covered
    4. Build — not covered
    5. Run — covered
    6. Improve — covered

    Picks up after launch: the run and improve gates, reviewed quarterly.

    Enquire about a retainer engagement
  5. 05

    Enterprise

    A multi-workstream programme with a dedicated team, governance and agreed service levels.

    Typical length
    6–18 months
    Pricing model
    Programme fee · by statement of work
    Best for
    Large organisations running change across markets, portfolios or business units

    Stages covered

    1. Discover — covered
    2. Define — covered
    3. Design — covered
    4. Build — covered
    5. Run — covered
    6. Improve — covered

    Every gate, per workstream, under a steering group and SLAs.

    Enquire about an enterprise engagement
  6. 06

    Squad

    A dedicated team that works inside your stack and sprint schedule.

    Typical length
    Ongoing · 3-month minimum
    Pricing model
    Time & materials
    Best for
    Teams with a clear roadmap that need more senior people quickly

    Stages covered

    1. Discover — not covered
    2. Define — not covered
    3. Design — covered
    4. Build — covered
    5. Run — covered
    6. Improve — covered

    Works to your team’s schedule; your leads own each gate and our team meets its criteria.

    Enquire about a squad engagement

Tell us what you need built.

You will speak to a lead who would run the work, and get a straight answer on fit.

Book a call

Three ways to start

  1. 01About 2 minutes

    A quick question

    You get A reply from a named lead

  2. 02About 8 minutesRecommended

    A project brief

    You get Options and a first scope after one call

  3. 03About 15 minutes

    A formal RFQ or RFP

    You get Receipt confirmed and a named bid lead

Every engagement starts with a written scope and a quote agreed before work begins. How each package is priced

See the work Explore our services Industries we serve Meet the team Email us Read our insights