elwood-ai.
Boutique AI consultancy · Singapore / global

We put AI inside the boring middle of your business.

Audits, builds, advisory, training. Putting AI into the work that runs your business.

Start a project → See the work
§ I · Practice

Four services. One engagement model.

We take on the whole arc of putting AI into a working process — from the first diagnostic, through building, to the training your team needs to own it after we leave. The audit is where we always start.

01 · Discovery

AI workflow audits

Where AI earns its keep inside a real process — mapped, ranked, honest.

Read more ↓
Fixed fee · 2–4 weeks

Two to four weeks. We shadow a real process end to end — intake, decision, handoff, exception — and return a mapped workflow with the three to five places AI will meaningfully change the unit economics.

Each play is ranked by cost and risk. The audit earns the right to build — or to walk away.

02 · Build

Custom agents & integrations

Copilots and drafting agents, retrieval over your own documents — shipped with tests and rollback.

Read more ↓
Scoped sprints · 6–14 weeks

Six to fourteen weeks of scoped sprints. Real users in the loop by week two. A staging environment your team can poke at.

Internal copilots, drafting agents, retrieval over your documents — inside the tools your team already uses, never a separate portal. Nothing goes live without a test, versioning, and a rollback.

03 · Advisory

Strategy & governance

Build-vs-buy, vendor diligence, policy — defensible in front of a board or regulator.

Read more ↓
Retainer or discrete reviews

For leaders who have to defend a decision in front of a board, a regulator, or an audit committee. Written up, with sources.

Build-vs-buy, vendor diligence, model and data policy, risk registers, procurement reviews, incident playbooks. We say when the honest answer is we don't know yet.

04 · Enable

Training for your team

Practical fluency for the non-technical people whose work just changed.

Read more ↓
Half-day to multi-week

Not for engineers. For underwriters, lawyers, analysts, salespeople — the hundred users whose work just changed.

Not prompt engineering. Practical fluency: how to use the system well, when to trust it, when to escalate. Half-day workshops to multi-week programmes, paired with a written playbook.

§ II · Vision
Most companies don't need a moonshot. They need the boring middle of their operation to stop costing them a week per week.

The model is the easy part. The hard part is the queue it feeds into, the reviewer who has to trust it, the regulator who asks how it decides, and the person whose job just changed.

So we work with four clients at a time, staff with senior people, and measure ourselves on whether the system is still trusted a year after we've left.

§ III · Selected work

Four projects, four industries.

A representative slice. Clients named on request, under NDA. Qualitative outcomes only — figures shared privately.

W-01

Underwriting intake triage

Specialty insurer · London

Broker submissions re-keyed by hand into the policy admin system. A grounded extraction layer with a reviewer queue.

Cycle time cut to hours.
Read case ↓

We built a grounded extraction and classification layer with a reviewer queue. Every submission routed by line of business, with field-level confidence shown to the underwriter.

Nothing booked without a human sign-off. Full audit trail from source document to policy record.

W-02

Field-sales co-pilot

Industrial distributor · global fleet

Reps on the road needed quoting and compatibility answers. Retrieval over the SKU catalogue, inside their mobile app.

Attach rate on complex orders, materially up.
Read case ↓

Retrieval over the SKU catalogue and historical orders, delivered inside the existing mobile app. Offline-tolerant for warehouse floors with patchy signal.

Research time per call way down. Attach rate on complex orders materially up.

W-03

Precedent & citation retrieval

Corporate law firm · mid-market

Citation-grounded retrieval over a decade of internal precedent. Every answer links back to the source document.

Hours returned to every fee-earner.
Read case ↓

Citation-grounded retrieval over a decade of precedent. Every answer links back to its source document — nothing synthetic, nothing invented.

Reviewed with the firm's librarians before release. Every retrieval logged for professional-privilege audit.

W-04

RFP response engine

Enterprise SaaS · multi-region

A drafting agent reading the approved answer library, with versioning and reviewer sign-off baked in.

First-draft time cut by most of a week.
Read case ↓

First-draft RFPs taking over a week. Four reviewers. No shared memory between responses — the same question answered three different ways in three months.

Reviewers still own the response. The agent assembles the first draft, then disappears.

§ IV · Commission a project

Tell us what you're trying to change.

A paragraph is enough to start. A real person reads every enquiry and writes back, usually within two working days.

Dear elwood-ai,

Tell us about the process — who touches it, where it breaks, what you've tried. Technical readiness comes later.

A partner replies within two working days. Nothing you send is used for training.

— the practice
Routes to info@elwood-ai.com