AI agent development: bespoke agents that read, match and prepare work for your approval

Our last agent cut a purchase-order task from up to 48 hours to around two minutes, after another agency handed the brief back. Built in three weeks, with a person approving every result.

Show us the process → Call 0115 647 0444

At a glance

  • 5.0 from 50 Google reviews
  • Based in Long Eaton, working UK-wide
  • East Midlands Chamber member

Vimana Digital builds bespoke AI agents for UK businesses: software given a goal, access to your documents and systems, and permission to work through the steps, then stop for a person to approve. Typical jobs are purchase orders, invoices, reconciliation and document intake. Our Webro agent was delivered in three weeks. Quotes are fixed-scope.

EmailAI readsQuoteApproved

What is an AI agent, in plain English?

An AI agent is software you brief like a capable new member of staff: here's the goal and the tools you may use, work out the steps, and check with me before anything leaves the building. Unlike an automation, it copes with variation. It reads "2-core 1.5mm flex, black, 100m drum", works out which of your products that is, and flags the lines it isn't sure about.

The word is being abused. Gartner (June 2025) estimated that only around 130 of the thousands of vendors calling their products "agentic" were building real agents, and called the rest "agent washing". Four questions sort the real from the relabelled:

  • Does it take actions, or only talk?A real agent opens the attachment, queries the catalogue, writes the draft.
  • Does it handle unseen inputs,or break when a supplier changes an invoice layout?
  • Is there a list of what it may do,and a log of what it did?
  • Where does a person step in?"Nowhere" means it's not an agent, or not safe.

The one-minute comparison of automation, chatbot and agent is on our AI services page.

What did the Webro purchase-order agent actually do?

The problem. Webro Cables & Connectors' customers send purchase orders as PDFs, sometimes with hundreds of lines, using their own part numbers. The only link to Webro's references is the product description. Someone had to read every line, find the Webro equivalent and build the quote: up to 48 hours for a big order.

  1. 1Receive email:A watched mailbox spots the purchase order.
  2. 2Extract PDF:The attachment is pulled into the workflow.
  3. 3Read items:Lines, quantities and descriptions become a structured list.
  4. 4Match parts:The agent reconciles each description against Webro's product data. This stage needed an agent rather than rules.
  5. 5Prepare quote:Matched lines become a quotation.
  6. 6Review and action:A person at Webro checks and sends it. Nothing reaches the customer without that step.

What are AI agents actually good at in a business?

Good at, right now:

  • Document intakepurchase orders, invoices and forms in inconsistent formats.
  • Matching and reconciliationyour references against theirs, bank lines against invoices.
  • Triage and routingwhich team a document belongs to, and why.
  • Drafting for approvalquotes, replies and summaries a person checks and sends.
  • Cross-system data entrywhere no connector exists, including software with no API, under tight limits.

What we'd talk you out of:

  • Payments or refunds without approval
  • Final outputs that need legal certainty,such as contract terms or regulated advice.
  • Low-volume tasksIf it happens twice a month, a person is cheaper.
  • Replacing staff on day oneAgents work best taking the repetitive work off someone's desk, with that person reviewing the rest.

Got a process with a "someone has to read it" stage in the middle? That's a good agent project.

Get a quote

How do you stop an agent doing something wrong or expensive?

  • An allowed-actions listIt can draft a quote; it cannot email a customer or change a price. Anything not listed is impossible.
  • Approval stepsbefore anything that sends, pays or changes a record.
  • Spend budgetsper run and per day, so a runaway loop costs pounds, not thousands.
  • Confidence thresholdsuncertain lines are flagged for a person, not guessed.
  • A full audit logof what it read, decided and called, and what the reviewer changed.
  • Staged rolloutobserve, then suggest, then act within limits, each step your decision.
  • Testing on last quarter's real documentswith sending switched off.

Which decisions can an AI agent legally make on its own?

Since 5 February 2026, UK GDPR Articles 22A–22D (inserted by the Data (Use and Access) Act 2025) generally permit solely automated decisions with significant effects, provided the person is told, can obtain human intervention and can contest the decision. Special-category data such as health stays more restricted. The test is whether there is meaningful human involvement.

The ICO's draft guidance on automated decision-making (consultation closed 29 May 2026) is reported to define "meaningful": a reviewer with authority to change the outcome, who understands it and considers each one rather than rubber-stamping a batch. A tick-box doesn't count.

Who owns it

  • Controller and processor confusion across several suppliers' systems: we document who is responsible for which data before the build.You
  • Purpose creep: our agents get the fields the task needs and nothing else.You
  • Accidental inference of special-category data from free text: we restrict what's extracted and log it.You

Why do so many AI agent projects get cancelled?

Gartner (June 2025) forecasts that over 40% of agentic AI projects will be cancelled by 2027, for cost, unclear value and weak risk controls. RAND's root causes are on our AI services page; measured studies explain our method:

  • Brynjolfsson, Li and Raymond (NBER, 2023)5,179 support staff given an AI assistant resolved 14% more issues per hour, 34% for the least experienced, and barely more for the most experienced.
  • METR (July 2025)16 experienced developers using early-2025 AI tools were 19% slower while believing they were 20% faster.

The lesson: general AI tools give uncertain gains that people misjudge. A specific agent built around one process, measured before and after, produced the Webro result, so every project starts with a baseline and a single bottleneck.

Which AI models and frameworks do you use?

The model is a component, not the product. We pick the language model that suits the task and keep it swappable, because prices change every few months.

Model Context Protocol (MCP) is a shared plug standard for connecting agents to tools and data, governed by the Linux Foundation's Agentic AI Foundation since 9 December 2025. Connectors built to it work across vendors, so changing AI provider doesn't mean rebuilding integrations.

Running cost. A document agent handling hundreds of orders a month typically costs tens of pounds in model usage plus hosting. We pick the cheapest model that passes the tests.

Where it runs. Your cloud account or ours (AWS or Azure), in UK or EU regions where offered, built in the same stack as our bespoke software. To confirm: Amit to confirm the Webro stack, model provider and hosting region

How much does AI agent development cost in the UK, and how long does it take?

UK agencies published these 2026 ranges: £7,000–£25,000 for a starter agent over 10–14 weeks, £25,000–£75,000 over 16–20 weeks (Tinderhouse); single-workflow agents at £10,000–£30,000 (OpenKit). Ours: To confirm: Vimana to confirm.

For comparison, the Webro agent went from first meeting to working demonstration in three weeks, against the previous agency's three-month estimate. We quote a fixed price and timeline after reviewing the process, covering testing, guardrails and hand-over, not just the agent.

How we deliver an AI agent

Testing follows our quality assurance standard. Talking to customers? You probably want a chatbot. Predictable steps between systems? Automation is cheaper.

  1. 1Review the briefWe'll say whether it's an agent job, an automation, or not worth doing.
  2. 2Map the process and set the baseline:Time taken today, failure points, what "done" means.
  3. 3Fixed-scope quote,Including allowed actions and approval points.
  4. 4Build against real history,Sending switched off.
  5. 5Staged rollout:Observe, suggest, act within limits.
  6. 6Measure and hand over:Before-and-after numbers, documentation, audit log and a named contact.

Frequentlyaskedquestions

Not covered here? Call and ask. We reply within 24 hours.

0115 647 0444Get a quote
What is an AI agent in business terms?

Software given a goal, access to specific tools and data, and permission to work through the steps itself: open the email, read the PDF, match each line to your catalogue, prepare the quote, then stop and ask a person to approve. It handles variation that fixed rules can't, which is what separates it from automation.

How is an AI agent different from a chatbot or an automation?

An automation runs fixed rules. A chatbot holds a conversation. An agent reasons across steps and tools to reach an outcome, such as reconciling a customer's part numbers with yours. Most useful systems combine them: automation for intake, an agent for the hard middle, and a human for the decision.

How much does AI agent development cost in the UK?

UK agencies publish ranges of roughly £7,000–£30,000 for a single-workflow agent and far more for multi-agent systems, with timelines of 10–24 weeks. We scope to a fixed price after reviewing your process. Webro's agent was delivered in three weeks for less than the previous agency had charged.

Can an AI agent make decisions about customers on its own?

Legally, UK GDPR Articles 22A–22D, in force since 5 February 2026, allow some solely automated decisions with safeguards, but the ICO expects a reviewer with real authority who can override. Practically, we build a human approval step before any consequential action, which also catches the agent's occasional mistakes before a customer sees them.

What happens when the agent is unsure?

It stops and asks. Each output carries a confidence level; anything below the threshold we agree with you is flagged and routed to a named person rather than guessed. On a workflow like Webro's, that means uncertain lines are highlighted for the reviewer, who checks every quote before it's sent.

What is MCP and why does it matter to a business?

The Model Context Protocol is an open standard for connecting AI agents to tools and data, governed by the Linux Foundation's Agentic AI Foundation since December 2025 with OpenAI, Google, AWS and Microsoft involved. For you it means integrations built once work with agents from different vendors, so changing AI provider later doesn't mean rebuilding everything.

Can an agent work with software that has no API?

Often, yes. Where a system only has a screen, an agent can be given controlled access to operate it like a user, or we extract data through exports, reports and documents instead. Screen-driving is slower and needs tighter guardrails and more testing, so we assess it case by case and say if it isn't worth it.

How long does an AI agent project take?

Published UK timelines run 10–24 weeks. Ours depend on scope, but the Webro agent, which reads purchase orders with hundreds of lines and matches mismatched references, went from first meeting to demonstration in three weeks. Simpler document agents can be quicker; the time goes into testing on real history and the approval design.

Show us the process, not the technology

If someone on your team has to read, interpret and match before anything can happen, send us the task. We'll tell you whether an agent is the answer, what it would cost and how long it would take.

Author: Amit Saini, founder and lead developer, Vimana Digital. East Midlands Chamber member. Last reviewed: 9 October 2026

"Amit is very knowledgeable and has been very patient with us, responding quickly to any changes we have asked for (and there has been a lot of changes!)." Excel Process Systems Ltd, Google review
What do you need?

Your details are only used to reply to your enquiry.