00 AI agents
An AI agent reads what arrives, works out what it is, does the next step in your systems and hands anything unusual to a person. Where AI automation runs on a trigger over internal work, an agent talks to your customers or your staff across channels and decides which system to touch. We build them into the CRM, inbox, WhatsApp, team chat, shared drives and databases your team already runs. Scoped access, a designed failure route and monitoring come with the build.
01 What it means
An AI agent is software that reads what arrives, works out what it is, does the next step in your own systems and hands anything unusual to a person. A chatbot answers a question and stops. An agent also updates the record or routes the request to the right person.
Most businesses already have the information. It sits in inboxes, CRMs, chat tools, PDFs, spreadsheets, call notes, shared drives and databases. What is missing is something that understands what each item means and what should happen next, so a person spends the day reading and forwarding. An agent does that reading and forwarding inside the tools you already pay for, with limits on what it can touch.
Once an agent has settled in, you stop noticing it. The right fields appear in the CRM. The right person gets the right summary. The risky case gets escalated. The log shows what happened. Nobody has to ask which inbox has the truth.
02 What it can do
A customer writes in. The agent works out what they are asking and sends the reply from wording you have approved where that is safe, or drafts it for a person to send.
Approved wording and an exception queue
When a chat or message thread ends, the agent matches the contact, writes the lead and a short summary into the CRM and creates the follow-up task.
Your CRM stays the record
Enquiries from your website chat and WhatsApp reach one backend, sorted the same way, with consent recorded and the right person sent a short summary.
One log across every channel
When the agent is unsure or the case is sensitive, it stops and passes the whole conversation to a named person, so the customer never has to repeat themselves.
A named owner for every handover
Regulated and specialist work needs routing that follows your rules and a clear audit trail. We design for both from the first day.
Audit trail from day one
Answers that have to come from your own policies and files are an AI knowledge base, which has its own page. Reading paperwork, tidying records in batches, internal triage, filing and chasing are AI automation. If what you need is a whole app, or a CRM across several locations, rather than an agent inside your existing tools, see custom software. These are jobs we design to, not a list of claims: the case we document in full is below, and every case we can describe in detail is on the case studies page.
03 How we build
01 · Where the truth lives
Which system owns each field, which records can be trusted and where your staff decide things by hand today. Your own system stays in charge. The agent works around it.
02 · Access
Scoped tokens, OAuth where it fits, least privilege and a named owner for each one, so the agent never depends on a shared personal login.
03 · When it fails
Retries, fallbacks, checks on the data and a place for exceptions on every important route. Quiet failure is not acceptable in production.
04 · Watching
Run status, cost, failed records, speed and the business events that show whether the agent is actually saving time.
04 What we check
The CRM, database, inbox or document store still owns the record. The agent reads, classifies, suggests and routes around it. An AI step never becomes the secret source of truth.
Each token can read, write or delete only what the job needs, with a named owner and a renewal rule. No shared passwords, no personal logins.
Can we read and write the fields the business needs? Are webhooks available, or do we need controlled polling? Will the rate limits survive a normal day? We find out before the budget is committed.
Every agent eventually meets a timeout, a malformed record, a revoked token or a changed API response. We decide what happens before that day, and failed records can be replayed without duplicates.
Low confidence, sensitive output or anything that could affect a customer goes to a person with enough context to decide quickly.
We estimate volume before launch and track usage after it. A useful agent should not surprise the business with quiet per-message or per-document costs.
05 Where agents stop
When none of these apply, and you need a production result without hiring a full internal team first, that is the work we do: discovery, build, documentation, monitoring and a handover someone else can operate. If you are not sure which side of the line you are on, that is what our AI consultancy is for. We look at how the work runs and tell you straight.
06 Before we talk
01 · What arrives
The enquiries, messages or documents that keep coming in: the shared inbox, WhatsApp, the website form, a shared drive.
02 · What happens next
Who reads each one, what they look up, and where they type the answer or the details afterwards.
03 · What goes wrong
What gets answered late, typed twice or missed altogether, and what you would hand over first if you could.
That is enough for us to draw the path and tell you whether an agent is the right answer or something simpler would do. If you are not sure which jobs are worth handing over, start with AI consultancy.
07 How we quote
Two to four weeks
Six to ten weeks
Rolling monthly
Each build is quoted as a fixed figure after discovery. Usage fees, hosting and third-party platform costs are itemised separately before launch.
08 Handover
We do not start by asking which model to use. We start by asking where the source of truth lives, who can read it, who can write to it and what happens if the output is wrong. That keeps the work grounded in the business process instead of the demo.
For regulated teams, that often means mapping personal data, checking whether a record holds sensitive information and deciding what should be logged. Some agents can store full prompts and outputs. Others should store only a status, the source record ID and a short audit note.
We design the monitoring path before launch too. A failed run should show which system failed, which records are affected, whether a retry is safe and who owns the next action. That is the difference between a managed agent and a hidden script.
We can run the agent for you on managed servers in the EU. How we run what we build.
09 Where this goes deep
Once an agent is connected, the work moves to what runs behind it: the workflows, the knowledge layer and, when your tools stop fitting, the software itself.
Triage, filing, chasing and reporting that run across the tools you have connected, with a person approving anything that matters.
Read moreAssistants that answer only from your own documents and policies, with sources shown and out-of-scope questions refused.
Read moreCRMs, client portals and internal tools built around how your team works, for when an agent inside your existing tools is not enough.
Read moreThe front desk answered and the compliance chased, for UK letting agencies and property managers.
Read more10 Case
A UK clinic had a website chat widget and a WhatsApp number feeding two separate inboxes. Both now reach one backend that sorts each enquiry, records consent and sends the right person a short summary.
CaseHealthcare servicesUK · anonymised
How we unified website chat and WhatsApp enquiries into one backend, with consistent classification, consent capture, team routing and conversation history.
Channel routing case study70%
Enquiries triaged
Sorted before human review.
2
Channels unified
Website widget and WhatsApp routed into one backend.
Names redacted out of respect for our clients.
11 Questions
What an agent is, what it can touch, how we keep it safe and what it costs, answered the way we would on a call.
An AI agent is software that reads what arrives, works out what it is, does the next step in your own systems and hands anything unusual to a person. A chatbot answers a question and stops, while an agent also updates the record or routes the request to the right person. A typical one reads an enquiry from your website or WhatsApp, pulls out the useful details, replies from approved wording or drafts a reply for a person, writes the record into your CRM and flags anything it is unsure about. It has limits on what it can touch, a log of what it did and a person to escalate to.
Yes. We built this for a UK healthcare services clinic: website chat and WhatsApp now reach one backend, each enquiry is sorted by what it is, consent is recorded and the right person gets a short summary with the conversation attached (the channel routing case). Where a question is routine and safe, an agent can answer from wording you have approved. Where it is not, it hands over to a person, and every conversation is logged.
Yes, provided the CRM lets us write the fields we need, and we check that before any budget is committed. After a conversation, the agent matches the contact, writes the details and a short summary and creates the follow-up task. Your CRM stays the source of truth, and the log shows exactly what the agent changed.
They can, and we usually recommend it, because an agent that says what it is earns more trust than one that pretends. You decide the wording and the point at which a person takes over. We build the handoff so the customer does not have to repeat themselves, because the person receives the whole conversation.
Three ways. It is scoped to the tools and fields the job needs, and nothing more. It answers only from approved wording or your own documents, and holds anything it is not confident about for a person. Every action is logged, with alerts on failures, so a mistake is caught and traced quickly.
Through the connections each system officially supports: OAuth where the platform offers it, scoped tokens with a named owner and a renewal rule, and webhooks where the source system supports them reliably, with controlled polling where it does not. Common ones are CRM platforms, email, WhatsApp providers, team chat, shared drives, databases, spreadsheets, website forms, booking tools and custom APIs. What matters more than the name is whether the platform lets us read and write the fields you need and has rate limits that survive a normal day. We check that before any budget is committed, and we will say no if the platform is too closed or unstable. We do not build on screen scraping, shared personal logins or unsupported private endpoints.
We start with an audit of where your data lives and who can reach it, then design the agent so credentials stay protected, data movement is logged and sensitive content only reaches the systems that need it. Each token is scoped to what the job needs, with a named owner and a renewal rule. We also decide upfront what gets logged: for some agents the full prompts and outputs, for others only a status, the record ID and a short audit note.
Yes, if it is designed with data minimisation, lawful basis, retention, access control and deletion routes from the start. We map which personal data each route carries, keep it to what the task needs and document how long prompts, outputs and logs are kept. For higher-risk uses we recommend a DPIA before production. Built in the UK. Hosted on managed servers in the EU, under UK GDPR.
A focused proof of concept usually takes two to four weeks: one agent, one path, real test data and output you can see. Production agents with several systems, business rules, monitoring, error handling and documentation usually take six to ten weeks. We build in that order on purpose, so you see the first path working before you commit to the full scope.
It is quoted as a fixed figure after a discovery call, because it depends on how many systems we connect, how much each platform lets us do and how much judgement the route carries. We do not publish a rate card, because a number on a website is priced to the website, not to your process. A proof of concept takes two to four weeks and a production agent six to ten. Usage fees, hosting and third-party platform costs are itemised separately before launch, so nothing arrives as a surprise.
We track trigger volume, failed runs, retry count, processing time, cost and the business events that matter. Alerts go to the people who can fix the problem, and they say which system failed, which records are affected, whether a retry is safe and who owns the next action. Failed records can be replayed without creating duplicates.
Yes. Managed support is a rolling monthly engagement covering monitoring, minor fixes, dependency changes and a monthly review of failures or missed cases. When a platform changes its API, we handle the update. If you would rather run it yourself, the handover pack covers owners, credential scope, retention, failure routes and how changes are tested.
Yes, where the job is routing. One agent can take enquiries from your website chat and WhatsApp into a single backend, sort them the same way and send each to the right person, with every message logged (the channel routing case). If you need a full CRM across several locations, with a pipeline and permissions for each, that is a build on our custom software page.
12 Enquiry
Tell us what arrives, where it lands and what should happen next. We come back with first questions or a discovery call slot.