An AI agent is software that works through a task in steps: it reads the request, looks things up in your systems, picks the next step within your rules and prepares the result for a person to approve. With Praxen, a custom solution with agents starts at £6,500 (£4,550 with the launch offer) and takes from 4 weeks, and a pilot on one workflow costs £2,900 (£2,030) and takes 2–4 weeks. Prices exclude VAT.
Chatbot, automation and agent: the difference
The three terms are often used as if they meant the same thing. They describe different tools, and the choice affects both cost and risk. Most useful systems combine them: a chatbot can pass a request to an agent, and an agent can start a fixed automation once a person approves its draft.
| Type | What it does | Example | Who decides |
|---|---|---|---|
| Chatbot | Answers questions in a conversation, from your approved information | Website chat that explains delivery times and returns | The customer asks, and staff take over anything outside the answers |
| Automation | Runs fixed steps you define, so the same input always takes the same path | A web form creates a CRM contact and sends a welcome email | The rules, set once by you |
| AI agent | Works towards a goal over several steps and chooses which tool to use next, within limits | Reads a supplier email, checks prices against the contract and drafts a reply | A person approves before anything leaves the business |
How an agent works, step by step
- Trigger. Something starts the run: an email arrives, a form is submitted or a scheduled time comes round.
- Read. The agent reads the request and works out what is being asked.
- Look up. It checks your systems, such as the CRM, stock list, calendar or shared drive, with read-only access unless you approve more.
- Choose the next step. It picks the next action from a list you allow, and stops if none fits.
- Draft. It prepares the result: a reply, an updated record, a quote or a summary.
- Approve. A person checks the draft and approves, edits or rejects it.
- Act and log. Only then does the agent send or save, and every step is recorded with the time and the source it used.
Five examples for businesses
- Supplier queries. A supplier asks when an invoice will be paid. The agent finds the invoice in Xero or QuickBooks, checks its status and drafts a reply with the payment date for finance to approve.
- Quotes from enquiries. The agent reads an enquiry, takes prices from your price list, checks stock or lead times and drafts a quote for a salesperson to review.
- Complaint triage. The agent finds the order, sorts the complaint by type and urgency, drafts a reply with the remedy your policy allows and routes it to the right person.
- Contract renewals. At a point you choose, such as 60 days before renewal, the agent sends the account owner a summary of the contract and a draft renewal email.
- Stock reorders. The agent compares stock levels with recent sales and drafts purchase orders for the buyer to approve, with the reasoning for each line.
Checks, logs and approval points
Every agent we build has these safeguards. You decide where each limit sits.
- Least access. The agent reads only what the task needs and writes only to the fields you agree.
- Approval points. Anything that leaves the business, moves money or changes a customer record waits for a person.
- Hard limits. The agent cannot make payments, change bank details or offer discounts outside your rules.
- A confidence threshold. When the agent is unsure, it stops and passes the case to a person with what it found.
- Full logs. Every step is recorded with its input, its source and who approved it.
- A pause switch. A named person can stop the agent at any time.
- Data in your accounts. The agent runs on your own accounts through enterprise APIs, which do not train models on your data, and we sign a data processing agreement before launch.
When an agent is the wrong tool
An agent adds a language model’s judgement to a process, and that judgement needs testing, monitoring and approval screens. In these cases a simpler tool, or no tool, is the better choice.
- The steps never change. A fixed automation is cheaper and more predictable.
- The volume is low. A task that comes up a few times a month rarely repays the build.
- Each case needs professional judgement, such as legal advice, a hiring decision or a credit decision. An agent can prepare the file, and a person decides.
- The data is scattered or out of date. Fix the source first, or the agent will repeat the errors faster.
What AI agent development costs
Agent work comes in three formats. Standard prices are £2,900 for a pilot, from £6,500 for a custom solution and £350 a month for ongoing care; the table shows launch offer prices. All prices exclude VAT. AI usage is billed by the providers at cost on your own accounts, and we estimate it before launch. The main cost drivers are the number of systems the agent connects to, the number of approval points and how many exceptions the workflow has.
These are illustrative figures. Say a wholesaler receives 40 order emails a day, and each takes six minutes to read, check and confirm. That is four hours a day. If an agent prepares each confirmation and a person spends one minute approving it, the daily time falls to about 40 minutes. In the audit we use your own volumes.
How to start
- Pick one task with a clear input and a clear output, such as an order email and a confirmation.
- Collect around 50 past cases, including the awkward ones.
- Write down the rules, the limits and every point where a person must approve.
- Agree how you will judge it, such as how many drafts people approve unchanged and how long each case takes from start to finish.
- Run the agent in shadow mode first: it drafts, your team works as usual, and we compare the results.
- Go live with approval on every action. Remove an approval step only where the results show it is safe and you decide to.