Service 03 · Execution office · Stage 04–05 of DATS

We ship it
into production.

An embedded AI delivery function that sits inside your organisation. We build the placement, wire it into your stack, stand up the evaluation harness, move it past go-live, and graduate the whole thing into your team. Long embed. Senior only. Outcome-weighted.

Book a scoping call See the phasesAll services

01 · Why execution, not implementation

Implementation firms leave. We graduate.

Every systems integrator in the world can ship an AI pilot. The hard part isn't the ship — it's the year after the ship. Day 90, when eval breaks. Day 180, when drift shows up. Day 270, when someone asks who owns it and no-one has an answer.

The Execution Office is our answer. We embed as your AI delivery function — small, senior, long-tenured — and we stay until the capability lives inside your team. Not a body shop. Not a retainer. A graduation path.

A placement isn't in production until someone on your team can turn it off without calling us.

This is the only engagement we offer where the success metric is our own replaceability. If twelve months in, you still need us to keep the placement alive, we've failed — and we'll tell you so.

02 · Four phases

A 12-month shape.

Longer engagements add more placements into the line — the shape stays the same.

  1. Phase 01

    Build + integrate

    Stand up the first placement. Real data, real integration, real users — not a sandbox. Pair-build with your engineers to transfer context as we go.

    Months 01–04
  2. Phase 02

    Eval + harden

    Evaluation harness live before production cutover. Guardrails, incident tooling, observability, rollback. The things that make audit sleep at night.

    Months 03–06
  3. Phase 03

    Production + run

    Cutover. On-call rotation. Drift monitoring. Quarterly review board. A second placement enters build while the first is in run.

    Months 06–10
  4. Phase 04

    Graduate + exit

    Runbook handover. On-call fully in-house. Eval framework your team owns. Dilr.ai shifts to quarterly check-ins — or the next placement.

    Months 10–12

03 · Who we embed

Small squad. Senior only.

A typical Execution Office squad is three to six people. We don't staff juniors on your engagement and charge partner rates. Every embedded practitioner has shipped AI in production at enterprise scale.

  • Role 01

    Placement lead

    Owns the placement end-to-end. Your single point of accountability. Runs the weekly cadence with your sponsor and the quarterly review with your board.

  • Role 02

    AI engineer(s)

    1–3, depending on scope. Model selection, prompt engineering, eval harness, guardrails, retrieval, fine-tuning. Ships code you can read.

  • Role 03

    Platform engineer

    Integration with your stack. Identity, data pipelines, observability, CI/CD, cost controls. The part everyone forgets until month four.

  • Role 04

    Eval lead

    Owns the evaluation framework — offline, online, adversarial, and customer-facing. Keeps the harness alive past go-live.

  • Role 05

    Governance partner

    Part-time. Bridges the operating model into the live placement. Handles risk, audit, and the difficult conversations with compliance.

  • Role 06

    Your team

    We pair on every piece of the build. By month twelve, your engineers are committing unaided and your product owner runs the placement without us in the room.

04 · How it plays out

A regulated wholesale bank had two years of pilots.
We shipped the first in four months — and the second into the same chassis.

The pilots weren't bad. The chassis didn't exist. Four months to first production cutover, three more to the second placement riding the same eval harness and the same governance boards. By month twelve, the bank's own engineers were committing. Our on-call rotation was fully in-house. We moved to a quarterly check-in.

Composite · details anonymised · representative of engagements

05 · FAQ

Questions, answered.

Isn't this just staff augmentation?
No. Staff augmentation ends when the invoices do. We design the engagement around graduation — if you still need us after twelve months, something's gone wrong. We explicitly tie part of our fee to your team's independence.
Do you work with our existing vendors?
Often, yes. Some placements land best on OpenAI. Some on Anthropic. Some on open-source. Some on your existing ML platform. We're vendor-neutral; we pick the placement-appropriate stack and we work alongside whoever else is in the room.
What if we don't have a placement identified yet?
Then start with Placement Diagnostic. Execution Office only makes sense once the placement is chosen and the operating model exists. We will push back on skipping those.
How many placements can you run in parallel?
The usual pattern is one live in run while the second is in build and the third is being scoped. More than three parallel placements requires a larger embedded squad or a second Execution Office lane.
How do you price outcomes?
A base retainer covers the squad. An outcome layer ties to specific, measurable placement KPIs — agreed up front, tracked on the eval harness, paid quarterly. We don't invent metrics; we use the ones already on your scorecards.
Who owns the IP?
You do. Everything we build inside your stack — code, prompts, eval harnesses, runbooks — is your IP, delivered in your repos. We retain rights only to generic methodology and tooling we bring with us.

Have a placement? Ready to ship it?

30-min scoping call · Senior-only embedded delivery · Outcome-weighted fees.

Book a scoping call See all services

From the blog

Field notes from the front line.

All articles