AI agents for business
We build, deploy and supervise AI agents that process documents, answer customers and draft quotations inside the systems you already run.
- First agent live
- 2-4 weeks
- Runs on
- Cloud or on-prem
- Human approval
- On by default
- Usual first process
- Document intake
What an AI agent actually is
An AI agent is a program with a goal, a set of tools and permission to act. It reads an incoming email or document, decides what needs doing, uses your systems to do it, and escalates anything it is not sure about to a person. The useful mental model is not a smarter search box. It is a new colleague who has been given access to a few systems and a clear set of instructions.
Technically every agent we build has the same three parts. A language model that reads text and makes the judgement calls. Tools, which are the connections into your email, ERP, CRM and file server. And rules, which define what it may finish on its own and where it has to stop and ask a human. The last part is the one that decides whether the project succeeds.
Agent, chatbot or RPA: which one is it
| Chatbot | RPA script | AI agent | |
|---|---|---|---|
| Handles free-form text | Yes | No | Yes |
| Acts inside your systems | No | Yes | Yes |
| Survives a changed layout | n/a | No, it breaks | Usually |
| Handles exceptions | No | No | Flags them for a person |
| Explains what it did | No | Log only | Reasoned, step by step |
| Good fit for | FAQ deflection | Fixed, rigid workflows | Messy input, clear rules |
The four agents companies actually buy
1. Document processing agents
Invoices, orders, delivery notes, contracts. A large share of administrative work is moving data from one format into another. The agent reads the incoming PDF or email, extracts the fields, checks them against your rules and writes them into the accounting or ERP system. Anything that does not reconcile, such as a total that disagrees with the order, goes to a review queue instead of straight through.
2. Customer support agents
Not a bot that offers links to a help page. An agent that can see order status, stock levels and delivery data, so when someone asks where their order is, it answers with the real number. Harder or emotionally loaded messages it recognises and routes to a person, with a summary of the situation already written.
3. Quotation agents
In manufacturing and trade, assembling a quotation takes hours: gathering prices, checking specifications, formatting the document. The agent pulls the parameters out of the enquiry, applies your pricing logic and produces a draft. A person reviews, adjusts and sends. Two hours of assembly becomes ten minutes of review.
4. Internal knowledge agents
Every company keeps knowledge in people's heads: how that form is filled in, who approves a discount, what to do when the machine throws that error. An agent connected to your internal documents answers those questions with a link to the source. For a new hire it removes weeks of onboarding. For everyone else it removes the interruptions.
How we deploy one
- Scope one process. One process that repeats daily and has rules a person can say out loud. Never a chaotic one: AI only speeds up chaos.
- Wire the tools. Read and write access to the systems the process touches, with the narrowest permissions that still let the job finish.
- Run in suggestion mode. The agent prepares, a person approves. Every step is logged, so you can see exactly what it read and why it decided what it did.
- Measure at 4-6 weeks. Hours saved, error rate, share of cases that needed a human. Real numbers, against the baseline we took before starting.
- Promote or stop. Steps where approval has become a formality get automated. Steps that keep needing judgement stay with a person, permanently and on purpose.
Guardrails, because this is the part that goes wrong
- Grounded answers only. The agent answers from your data. If the question falls outside it, it says so and hands over rather than inventing a plausible reply.
- Hard stops on money and commitments. Payments, discounts and contract terms always require human approval. No exceptions, however good the statistics look.
- Full audit trail. Every input, decision and action is logged. When something is wrong six weeks later, you can see precisely where it went wrong.
- Scoped access. The agent gets the minimum permissions the task needs. It cannot reach systems that are not part of its job.
When an AI agent is the wrong answer
We will tell you this on the call rather than after the invoice. An agent is a poor fit when the process runs a handful of times a month, when nobody can explain the rules consistently, when the source data lives only on paper, or when a deterministic script would do the same job cheaper and never surprise you. Three or four times a year the honest answer is a database view and a scheduled email, and that is what we say.
What drives the cost
Three things, in this order: how many of your systems the agent has to touch, how clean the source data already is, and whether it runs in the cloud or on hardware you own. The model itself is rarely the expensive part. A pilot confined to one process is the smallest useful unit of work and the cheapest way to get a real number for your case, which is why we recommend starting there rather than with a company-wide programme.
If you do not yet know which process to point it at, that is what the AI readiness audit produces in a week. If the documents in question cannot legally leave your network, the same agent can run on your own servers.
Questions we get asked every time.
What is the difference between an AI agent and a chatbot?
A chatbot answers a question and stops. An AI agent has a goal, tools and permission to act: it can open an email, read the attached PDF, look the customer up in your CRM, write the data in and notify the person responsible. A chatbot talks about the work. An agent does the work.
How long does it take to deploy an AI agent?
A first agent in one well-understood process takes two to four weeks from the go-ahead. Most of that is not the model, it is the plumbing: access to the source system, the exception rules and the approval screen. Agents in processes with messy or undocumented rules take longer, which is what the audit is for.
Will an AI agent make mistakes?
Yes. Less often than a tired person doing repetitive work, but it will. That is why we run every new agent in suggestion mode first: it prepares the action and a person approves it with one click. Once the approval statistics show the reviews have become a formality, specific steps get promoted to full automation.
Can an AI agent work inside our existing systems?
That is the point of one. Agents connect through the APIs your ERP, CRM, email and file storage already expose, so nobody changes tools. Where a system has no API, we integrate through database access, file drops or a scripted interface. Older on-premises software is normal and rarely the blocker people expect.
Who maintains the agent after it goes live?
We do, under an SLA, unless you would rather own it. Agents need maintenance for the same reason employees need updated instructions: suppliers change document formats, processes change, new exceptions appear. An agent nobody reviews degrades quietly over about a year, and the failure is hard to notice.
Keep reading.
Which repetitive process should the first agent take?
Bring one process that repeats daily and eats hours of somebody's week. Thirty minutes and we will tell you whether an agent can take it.
Book a free 30-minute call