Service detail

Last updated: September 2026

Automate one workflow, with review where it matters

Turn a repeated manual task into a small workflow your team can inspect. Start with representative examples and the rules for what a person must approve.

Quick DX PoC: £9,500-£13,500 · 2 weeks

01

Automate a task with a visible decision

A good first workflow has a recognizable input and one useful output. For example, extract agreed fields from an incoming document and show them beside the source before a reviewer approves them.

  • Choose one document family or task, with clear supported inputs.
  • Keep proposed values and source information visible together.
  • Route missing, uncertain or unsupported results to a person before downstream use.
02

What your team receives

The deliverable is the agreed input-to-review path, together with the evidence and instructions needed to judge it. Model choice follows the task and access constraints.

  • Input handling and field rules for the agreed formats.
  • A review surface with the source, proposed result and correction path.
  • Explicit handling of empty responses, missing fields and dependency failures.
  • An evaluation record, limitations and setup notes tied to the reviewed version.
03

Agree how to judge useful output

A demo using one clean document is insufficient to define acceptance. Choose representative cases, define what a reviewer considers correct and agree any thresholds before implementation.

  • Keep evaluation examples separate from the examples used to develop the workflow.
  • Include incomplete, ambiguous and unsupported documents in the agreed checks.
  • Verify that corrections can be made before a proposal is used.
  • Record a provider failure as a failure, preserve useful input and expose the agreed recovery path.
04

Start with examples and a reviewer

Bring representative files, the fields you need, a description of the manual task and the person who can evaluate the output. Use an approved sharing channel for the data.

  • Identify which fields can be optional and which must block progress.
  • State what the system may suggest and what a person must approve.
  • List formats, languages, volumes and any access constraints as facts or open questions.
  • Establish a manual baseline before estimating time savings or business impact.
05

Keep the first scope explicit

The package below is a starting point. Inputs, exclusions, timing and price are agreed in writing before work starts. A review workflow and an unattended operation require different acceptance conditions.

  • Automatic external messages, unattended final decisions and additional document types are excluded unless agreed.
  • Production integrations, a complete knowledge-base migration and ongoing model monitoring are separate scope decisions.
  • Real-data permissions, provider costs and operational responsibilities are confirmed for the engagement.
06

Use the result to decide the next investment

The final review compares the output with the agreed checks and makes unresolved limitations visible. The next step may be data preparation, a narrower workflow or a separately scoped rollout.

  • Review successful examples alongside failures and untested conditions.
  • Assign owners to corrections, operating tasks and any follow-up evaluation.
  • Retain the reviewed configuration, fixtures and setup notes so the team can repeat the check.

Delivery examples

Inspect the deliverables before you commit

Filled-in examples show how the scope and review evidence are documented.

Buyer FAQs

Does every workflow need an LLM?

No. Rules, validation or a simpler integration may solve the task more predictably. We assess that before making a model part of the delivery.

Can you guarantee an accuracy percentage?

We agree the evaluation set, scoring method and acceptance threshold for the actual task. A general accuracy promise without those definitions would not tell your team how the workflow performs.

Will the workflow act without approval?

Only actions explicitly included in the scope are implemented. The example on this page keeps a person responsible for reviewing proposals before downstream use.

What if our sample data is inconsistent?

That becomes a scoping finding. Clarifying fields, labels or input quality may be the right first step before funding an automated build.

Scope the first sprint

Bring the app, API, LLM feature, or AI workflow you want to test. We will turn it into a clear first-sprint scope.

Start a conversation