Comtau builds AI applications, agents, automations and assistants for companies that need AI inside real processes - connected to their data and systems, tested on their own cases, and reviewed by people where decisions matter.
01 / What this service is
An AI feature that impresses in a demo and an AI application your team relies on every day are different projects. The second needs data connections, permissions, tests, monitoring and a plan for when the model is wrong. That second kind is what our AI development services deliver.
Definition · AI application development
Designing and building software in which an AI model performs part of a business task - reading, classifying, drafting or deciding the next step - together with the interface, integrations, controls and tests that make it dependable in daily use.
Teams that read, sort, re-key and chase the same kinds of documents and requests every day, and want that work handled with fewer manual steps.
AI product development for companies adding AI features to a product they sell - search, summaries, assistants, automation - that have to work for every customer, not one demo account.
Owners of a CRM, ERP or internal platform who want AI inside it without rebuilding it, and without data going where it shouldn't.
Teams with a list of AI ideas and no clear first step. For that stage, AI consulting turns the list into a prioritised plan.
02 / What we build
Our AI application development services cover six kinds of work. Pick the one closest to your project; each service page explains the architecture, controls and process for that kind of system in detail.
Software that completes a defined multi-step job - reading records, calling tools, preparing changes - inside the permissions you set, and stops for approval where the rules say so.
AI agent developmentBusiness processes where AI reads and sorts the messy inputs, rules check them, and people handle only the exceptions.
AI workflow automationDrafting, summarising, extracting and searching, grounded in your documents and checked before anything is sent.
Generative AI developmentConversational help for customers and staff that answers from approved sources and hands over to people cleanly.
AI chatbot developmentLanguage-model features added to software you already run, with your access rules and a rollout you can switch off.
AI integration servicesA scoped assessment of where AI is worth using in your company: opportunities ranked by value and feasibility, data readiness, build-or-buy options and a first project defined well enough to start.
AI consulting servicesEach chain shows a service's typical mechanism in the notation used across this site. Diagrams use fictional data and are not client projects or Comtau products.
03 / How it fits together
The model is one layer of five. Most of the engineering - and most of the reliability - sits in the layers around it, and in the checks that run alongside all of them.
Where people use the result: a screen in your application, a chat, an email draft, an approval queue. Designed so users can see sources and correct the output.
The code that decides which steps run, in what order, with which inputs - and which rules apply before anything is saved or sent.
A language or other AI model doing the part that needs judgement: reading, classifying, drafting, deciding the next step. Replaceable behind one interface.
Search across approved documents and records at request time, filtered by what the user is allowed to see, so answers rest on your facts.
Connections to CRM, ERP, helpdesk, document storage and internal APIs, with read and write permissions set per action.
Running alongside every layer
Input and output checks, permissions, approval gates, limits on spend.
A test set of real cases run before every release and every model change.
Logs, traces, quality and cost per request in production.
04 / Use cases
Examples of tasks AI applications handle well today. The last column matters as much as the first: it is where a person stays responsible.
| Function | What the AI does | Systems involved | Where a person decides |
|---|---|---|---|
| Customer service | Answers questions from policies and order data, drafts replies, summarises long tickets for the next agent. | Helpdesk, order system, knowledge base | Refunds, exceptions, complaints |
| Operations | Reads incoming documents and requests, extracts the fields, routes each item to the right queue. | Email, document storage, ERP | Low-confidence items and approvals |
| Finance | Matches invoices to orders, flags mismatches, prepares reconciliation notes and month-end summaries. | ERP or accounting, banking exports | Payments and postings above limits |
| Sales | Prepares account briefs, drafts proposals from approved content, updates CRM records after calls. | CRM, proposal library, calendar | Pricing and commitments |
| Legal & compliance | Checks contracts against required clauses, compares versions, collects evidence for audits. | Document management, contract repository | Every legal judgement |
| Product | Adds search, summaries, assistants or classification to the software you sell. | Your application and its data | Release decisions, quality thresholds |
05 / Choosing an approach
Part of AI software development is deciding where a model is not needed. Fixed rules are cheaper, faster and fully predictable; a model earns its place only where inputs vary or judgement is required.
06 / Risks and controls
Each risk below is designed for from the first prototype, not added after an incident.
07 / How an engagement runs
Custom AI development is staged so you decide on each next stage after seeing the result of the previous one. The prototype stage exists to find out cheaply whether the approach works on your data. If you already have an AI prototype that must reach production, see taking an AI MVP to production.
The task, the people involved, the systems and the risk. An honest view on whether AI is the right tool.
out →Go / no-go and next step
Sample documents and records, integration points, quality bar and approval rules are written down.
out →Scope, evaluation criteria, estimate
The core AI step runs on your real examples and is scored against an agreed test set.
out →Test results, cost per task
Interface, integrations, permissions, monitoring and review queues, in short cycles with demos.
out →Tested release candidate
One team uses it under supervision; findings are fixed before wider rollout and handover.
out →Live system, documentation
08 / Where it runs
Enterprise AI solutions are judged on more than answer quality. We design for the requirements your security and IT teams will ask about:
Many AI applications reach people through an app they already use. AI app development covers the parts users touch as well as the services behind them:
09 / Budget
Two projects with the same label can differ by an order of magnitude in effort. An estimate becomes meaningful once the task, the data and the quality bar are written down - which is what discovery is for.
We don't publish prices that would ignore these differences.
Drafting for review is simpler to build and test than acting alone. Autonomy adds controls and test scenarios.
Clean, accessible sources make retrieval straightforward; scattered or scanned material needs work first.
Each connected system, and the quality of its API, often drives effort more than the AI step itself.
Higher accuracy targets need larger test sets and review flows; volume decides the running cost per request.
10 / Choosing a partner
Any artificial intelligence development services provider can show a convincing demo. These five questions separate a demo from software your team can depend on - ask them of every AI software development company you consider, including us.
11 / Questions
It depends on four things more than on the "AI" itself: how much of the work the system has to do on its own, how much data preparation is needed, how many systems it connects to, and how high the quality bar is. We do not quote a figure before seeing the task. A short discovery with sample data produces a written scope and an estimate you can compare with other options.
Plan for running costs as well: model usage is billed per request, so we estimate the cost per transaction during the prototype, before anything is built at scale.
The first useful milestone is a prototype tested against your own examples, which shows whether the approach works before larger investment. The full timeline depends on integrations, review workflows and how the rollout is staged. We give a timeline after discovery, not before, because the data and the systems involved decide it.
We are not tied to one provider. Depending on quality, cost, latency and where your data may go, a project can use hosted models from the major providers, models hosted in your own cloud account, or open-weight models run on your infrastructure. The application is built so the model can be changed later without rewriting it, and the choice is tested against your evaluation set rather than made on reputation.
Not by us, and the architecture is designed so it does not have to be. Most business applications retrieve your data at request time instead of training a model on it. We review each model provider's data terms with you and can choose deployments where requests are not retained or used for training, or keep processing inside your own environment.
Yes - that is usually where the value is. During discovery we list every system involved, what the AI needs to read, what it may write, and how access is controlled. Integrations use your systems' APIs with the narrowest permissions that do the job. If AI should live inside software you already run, see AI and LLM integration.
Buy when a packaged tool already covers the task and your data can go where the tool needs it. Build when the process is specific to how you operate, several systems must work together, or the AI is part of what you sell. Many projects combine both. If you are not sure yet, AI consulting is designed to answer exactly this question.
Rarely at the start. Most business tasks are solved by a capable existing model plus retrieval of your own documents and clear instructions. Fine-tuning becomes worth considering when you need a consistent format or style at high volume, or a smaller, cheaper model for a narrow task - and only once an evaluation set shows the simpler approach is not enough.
Working software deployed to the agreed environment, the source code, the evaluation set and test results, configuration for prompts and models, monitoring, and documentation your team can operate from. Ownership and support terms are set out in the contract before work starts.
Start node · your brief
Tell us what happens today, which data and systems are involved, and where a person must stay in control. We'll suggest a sensible first step - even if it isn't a build.
