N° 03 / Services — AI Agents

Most AI agents fail because nobody mapped the workflow first.

An agent is only as good as the process underneath it. We start by mapping the workflow — where judgement is genuinely mechanical, where it isn't, and where a human still needs to sign. The model comes last.

We've built document-extraction agents, internal copilots, and triage tools for teams in compliance, pensions, and operations. Each one was tested against the real data and the real edge cases, not a curated demo set.

How we engage

Three shapes of engagement. Pricing is per engagement — the discovery sprint is fixed-fee and a low-risk way to begin.

Discovery sprint2–3 weeks · Fixed fee: We map the workflow, identify the mechanical judgement, and tell you honestly whether an agent earns its place. You get a written recommendation either way.
Embedded build6–12 weeks · Quoted per engagement: We build the agent, the evaluation harness, and the guardrails alongside your team. We ship to production and stay until it holds.
Advisory retainerOngoing, monthly · Monthly retainer: For teams running agents already. We review evaluations, tighten guardrails, and keep the system honest as the work changes.

When agents make sense

Makes sense when

The workflow is already understood and repeatable
The judgement involved is mechanical, not political
There is real data to test against
A human can still sign off where the stakes require it

Step back when

The process itself is broken — fix that first
The decision needs accountability a model can't hold
AI is the goal rather than the outcome
There is no way to measure whether it worked

How an agent fits

An agent should slot into the workflow that already works — taking the mechanical load and leaving the judgement to a person.

Fig. 01 / Agent in an existing workflow

Existing workflow

The process as the team runs it today.

Mechanical judgement

The lookups and checks that don't need a person.

Agent

Runs the mechanical part, inside its guardrails.

Human sign-off

Judgement and accountability stay with a person.

source · kolasolutions.co.ukmonochrome · accent marks the agent

In practice

A few projects in this lane. Each one is a real engagement with a full case study.

ClassPro: A class-management platform with AI woven through it — workflow builder, report cards, year-wrapped summaries — for music schools, dance studios, gyms, and the long tail of class-based businesses.
Project SAM: An agent that takes the inbox as a job — filtering, summarising, drafting, booking, and chasing quotes — so a business owner reads three lines a day.

Frequently asked

Will an agent replace my team?: No. The agents we build remove mechanical keying and triage so the team spends its judgement where judgement is needed. The work tends to move, not disappear.
Which model do you use?: Whichever fits the task and the constraints. The model is an implementation detail; the workflow and the evaluation harness are the work.
How do you know the agent is working?: We build an evaluation harness before we build the agent. If we can't measure whether it's right, we don't ship it.
What about data leaving our environment?: We design for your compliance posture from the first conversation — private endpoints, on-premise, or redaction at the boundary. We've built for regulated teams; this is normal.
What if the process changes?: It will. The guardrails and evaluations are built so a change surfaces as a failing test, not a silent wrong answer. The advisory retainer exists for exactly this.
Can you start small?: Yes — the discovery sprint is designed to. Two to three weeks, a fixed fee, and an honest recommendation at the end.

N° 07 / Contact

Tell us what you're trying to build.

A short brief is enough to start. We read every enquiry ourselves and reply within two working days.

Coordinates

Kola Solutions Limited
Office 1426, 3 Fitzroy Place
1/1 Sauchiehall Street
Finnieston, Glasgow
Scotland G3 7RH
55.8655° N · 4.2809° W

[email protected]
Company no. SC781583