Confident output with no check
A model will return a tidy invoice total or category even when it has guessed. Without validation against your own records, the error travels downstream unnoticed.
AI workflow automation
AI Workflow Automation with People Still in Charge
We add AI to the steps in your existing processes that rules cannot handle — reading a messy email, pulling fields from a document, deciding which queue a request belongs in, drafting the reply — and wrap each one in validation, confidence checks and an approval queue, so the work speeds up without anyone losing sight of it.
- Every AI output validated before it is used
- Uncertain items routed to a person, with reasons
- Runs inside the tools your team already uses
Why AI steps misfire
Dropping a model into a process is not the same as automating it.
A prompt that works on ten examples in a chat window meets thousands of real items in production — scanned at odd angles, written in a hurry, missing half the details. That gap is where most of the engineering goes.
The AI steps that do most of the work
Nearly every useful AI step in a business process is one of these, and each can be measured on its own:
- Classify: what is this, and how urgent is it?
- Extract: which names, dates, amounts and references does it contain?
- Route: which person, team or system should handle it next?
- Draft: what should the reply, note or summary say?
- Check: does it match the rules, the order or the contract?
Review everything, or nothing
Teams either check every AI output — so nothing gets faster — or none of them, so mistakes reach customers. The useful middle ground is routing by confidence and risk.
No plan for the odd ones
Handwritten forms, forwarded chains and attachments in the wrong format get stuck or silently dropped, because nobody designed where the exceptions go.
A cost per item nobody forecast
Sending whole email threads and long PDFs to the largest model for a simple classification makes each item far more expensive than it needs to be.
What we build
AI steps built to slot into real processes.
We do not replace your workflow tool or your team’s routine. We add AI where the work needs reading or writing, and engineer the parts around it — checks, queues, fallbacks and logs — that make it safe to rely on.
Classification and triage
Emails, tickets, forms and letters sorted by type, urgency and language using your own categories, with the reason for each label stored alongside it.
Discuss this
Field extraction with validation
Names, dates, amounts and references pulled into a fixed schema, then checked against your CRM, ERP or order system before anything is written back.
NLP and document processing
Routing and assignment rules
The AI suggests where each item belongs; plain rules you can read decide what happens next, based on category, customer, value and confidence.
Business process automation
Drafting replies and case notes
First drafts of customer replies, case notes and internal summaries, grounded in the item and the records it relates to, ready for a person to edit and send.
Grounding in your documents
Confidence-based review queues
Only the items that need a person — low confidence, high value or a failed check — reach the queue, shown with the source, the AI output and the reason it was held.
Review tools and internal apps
AI steps your workflow tools can call
Each AI step exposed as a small service that Power Automate, Zapier, Make, n8n or your own code can call, so you keep the automation tools you already run.
API and integration work
How we work
How we add AI to a running process.
The seven stages we use on every project, adapted for a process that already runs: we label real items first, measure each AI step on its own, and widen automation only as the results support it.
- 01
Discovery
We pull a sample of real items — emails, forms, documents — and trace each through the process with the people who handle it, noting every point where someone reads, decides or types.
A sample set and the steps worth giving to AI
- 02
Strategy
For each step we choose a rule, AI or a person, pick the smallest model that meets the accuracy the step needs, and agree the confidence and value thresholds that send an item to review.
A step-by-step design with thresholds
- 03
UX & architecture
We design the review queue around the reviewer: what they see first, how they correct the AI in one move, and how each correction is captured to improve the step.
Reviewer screens and a correction flow
- 04
Development
Two-week sprints. Each AI step is built with a structured output schema, validation against your records and a labelled test set drawn from your own items.
AI steps working on real samples
- 05
Testing
Each step is scored separately — classification, extraction, routing — and the whole flow is run on a held-back sample, including the awkward items: scans, forwards and mixed languages.
Accuracy per step, on your data
- 06
Launch
A shadow run first, with AI outputs recorded but not used. Then items above the agreed confidence go straight through, while the rest wait for a person.
A live flow with a review queue
- 07
Optimisation & support
We track correction rates, move thresholds only as accuracy is proven, adapt prompts to new kinds of item and watch the cost per item as volumes grow.
Regular accuracy and cost reporting
Technology
The tools behind our AI steps.
Small, testable services that sit alongside the systems and workflow tools you already use, with the model layer kept swappable.
Models
GPT, Claude and GeminiSmall, fast models for classificationOpen models in your own cloudOCR for scans and photos
Workflow tools we plug into
Microsoft Power AutomateZapier and Maken8nYour own queues and jobs
Structured output
JSON schemasPydantic and Zod validationLookups against your recordsRetries on invalid output
Where work arrives
Outlook and Gmail shared inboxesHelpdesks and CRMsWeb formsSharePoint and file drops
Human review
Review queuesTeams and Slack approvalsCorrection captureRole-based access
Measurement
Accuracy per stepCorrection and override ratesCost per itemAlerts on failed runs
If a step can be done reliably with a rule or a lookup, we use the rule. AI is kept for the steps that genuinely need reading or writing.
Use cases
AI steps in everyday UK operations.
Each of these keeps the existing process and the existing systems. AI takes on the reading and drafting; a person keeps the final say.
First notification of loss intake
Claim emails and forms classified, policy numbers and incident details extracted and checked against the policy system, and the claim opened in the right queue for a handler.
Spotting complaints on arrival
Messages that amount to a complaint are flagged even when the word is never used, so they follow your complaints process and its timelines rather than sitting in a general queue.
Remittance advice matching
Remittance emails and PDFs read for invoice references and amounts, so payments can be matched to open invoices and only the leftovers go to a person.
Delivery exception handling
Courier notifications and customer emails about failed or late deliveries sorted by cause, with a suggested next action and a drafted customer update for the team to approve.
Repair request triage
Tenant repair requests sorted by trade and urgency, matched to the property record and passed to the right contractor queue — with anything safety-related sent to a person straight away.
New matter set-up from emails
Details from instruction emails and client documents extracted into your practice management system or CRM as a draft record, with conflict checks left to the team.
Relevant work
Products we have designed and built.
A selection of client projects related to this work. Each case study covers the brief, the approach and the stack.
All case studiesWhy Techsleight
What working with us is actually like.
No inflated numbers — just how we run projects, and what you can hold us to.
Product engineering, not ticket-taking
We ask what the software is for before we estimate it — and we will tell you when something should not be built, or should be bought instead.
AI where it earns its place
LLM features, retrieval and automation built with evaluation, guardrails and cost controls, and plain software where that is the better answer.
Full-stack under one roof
Design, frontend, backend, mobile, cloud and QA in one team, so nothing falls between suppliers.
UK-focused delivery
UK business hours, estimates in pounds, and a contract with a UK company. Our engineers are based in the UK and India.
Flexible engagement
A fixed-scope project, dedicated developers or a monthly retainer — and you can move between them as the work changes.
You own everything
Code, IP, cloud accounts and documentation are yours from day one. We sign an NDA before discovery if you need one.
Support after launch
We stay on for fixes, upgrades and new features, or hand over cleanly to your in-house team with the documentation to match.
FAQs
Questions we get asked.
Straight answers on scope, cost, timelines and how we work. If yours is not here, ask us directly.
Which steps in a process are worth giving to AI?
The ones where a person reads or writes unstructured text: sorting a mixed inbox, pulling details out of documents, deciding which team should handle something, drafting a reply. Steps that follow clear rules — a threshold, a lookup, a date check — are cheaper and more reliable as plain code, so we keep them that way.
How do you decide which items a person needs to check?
By combining signals rather than trusting the model’s own opinion of itself: whether extracted fields match your records, whether the output passes validation, the value or risk of the item, and whether it looks like anything seen before. Thresholds start cautious, so more goes to review at first, and are relaxed only as measured accuracy supports it.
Can you tell us how accurate the AI steps will be before we start?
We cannot say before seeing your items, and we will not quote a figure in advance. During the pilot we label a sample of your real items and measure each step against it — classification, extraction and routing separately — so you see accuracy on your own data, and set the thresholds, before anything runs unattended.
Can you add AI to Power Automate, Zapier or n8n flows we already have?
Yes. We usually build each AI step as a small service your existing flow calls, which keeps validation, logging and model choice in one place you control. Where a platform’s own AI action is good enough for a low-risk step, we will say so and help you configure it instead.
What happens when the model provider has an outage?
Nothing is lost. Items wait in a queue and are retried, a second model can take over for steps where that has been tested, and anything time-sensitive is routed to people with an alert. The process slows down rather than stopping, and every item is accounted for once the provider recovers.
How is the running cost per item kept down?
By using the smallest model that meets each step’s accuracy target, sending only the part of the document the step needs, caching repeated work and batching items that are not urgent. We estimate the cost per item during discovery. Model fees are paid directly to the provider, so you see them on your own bill.
Can the AI make the final decision on a claim, refund or application?
Not in anything we build. The AI sorts, extracts, checks and drafts; a person approves outcomes that affect people, money, health or legal matters. That is our design rule, and it also keeps you on safer ground under UK GDPR’s rules on solely automated decisions. Your DPO or legal team confirms what each process needs.
Do reviewers’ corrections make the system better over time?
Yes, deliberately. Every correction is recorded with the original item. We use them to add examples to prompts, extend the test set and spot new categories as they appear, then re-measure before changing anything in production. Your data is not used to train the model provider’s models through their business APIs by default; we confirm the terms for your chosen provider.
How does this differ from your business process automation work?
Business process automation covers the whole flow — integrations, rules, approvals and before-and-after measurement. This service is about the AI steps inside such a flow and the engineering that makes them dependable. Most projects include both, and if a process involves no free text to read or write, it may not need AI at all.
What does a first AI workflow pilot cost?
A discovery sprint from £2,000 samples your items, picks the steps and ends with a fixed-price quote for a pilot on one process. Model usage is a separate running cost, paid directly to the provider, which we estimate per item before you commit to the build.
Start a project
Which step does your team read and retype all day?
Describe the process and a rough monthly volume. We will tell you which steps suit AI, which should stay as rules, and how we would measure the result.
What happens next
- 1A senior engineer reads your brief within one working day, and replies with questions or a first view.
- 2A 30-minute call to understand the goal, constraints and what good looks like — no sales script.
- 3A written proposal with scope, milestones, team and a GBP estimate you can take to your board.
Techsleight Labs is a trading name of Krapton IT Consultancy.
Explore