Services / AI Agents
Custom agents that log what they did
Two things decide whether an agent survives its first bad week: whether you can read back what it did, and whether it knows the point at which it must stop and fetch a person. Both are written into the build before the first line of script. Voice, chat, assist, ops, code, data, multi-agent, and lines that are not on the board yet. Netherlands and EU, English engagements.
Anatomy
What makes it an agent
Five parts, and a thing missing any of them is a demo. The last one is the part most projects skip, and it is the reason staff will or will not leave the system alone.
Input
A call, a chat, a schedule, a webhook, or a queue. Something has to trigger it, and that trigger has to be nameable.
Tools
The CRM, the phone line, the repository, the warehouse, the knowledge base. Whatever the job genuinely touches, with access limited to that job and nothing wider.
Guardrails
Written limits and confidence rules, so that the agent has somewhere it is not allowed to go and knows when it has arrived there.
Trail
A log a colleague can read next week without asking you what happened. If nobody can audit the run, nobody will trust it after the first surprise.
Escalate
Escalation means one thing here: a named condition on which the agent stops, and a named person it hands the job to. Every line on the board carries its own, because a receptionist and a code reviewer fail in completely different ways.
Sample lines on the board
34 sample kinds, and a custom line whenever the job you run is missing from them. Each card names the tools it holds and the condition on which it stops and asks for a person.
Custom Agent
Built for the exact job you run, including the kinds this board does not list.
Tools and trail
Any team with a workflow that repeats often enough for someone to have a routine for it.
- Your stack
- APIs
- Logs
- Name the job and success metric
- Map tools and write rules
- Ship with trail + escalate
- Tune on real exceptions
Escalation rules are written with you before the build, and anything outside the named job stops by default.
Voice Agent
Answers and places calls, then books, qualifies, and hands off.
Tools and trail
Practices and offices where the phone rings more often than the desk can answer it, especially after hours.
- Phone
- CRM
- Calendar
- Inbound answered
- Intent classified
- CRM write + book
- Human if unclear
Hands to a person the moment a caller asks for one, or when the request falls outside the script it was given.
Chat Agent
Web and WhatsApp conversations that write back to the CRM instead of sitting there as FAQ wallpaper.
Tools and trail
Sites that already get traffic but have nothing live on them to qualify the people who arrive.
- Web
- CRM
- Visitor opens chat
- Qualify in a few turns
- CRM + booking
- Human if legal
Passes the visitor to a human on any question it has already failed once, and on anything commercial it was not given a number for.
Personal Assistant
Calendar, briefs, follow-ups for a founder or lean team.
Tools and trail
Founders and lean teams losing whole afternoons to preparation and chasing before any of the actual work starts.
- Calendar
- Docs
- Day loaded
- Briefs drafted
- Follow-ups queued
- Human for sensitive
Drafts but never sends on your behalf to anyone outside the company without your approval.
AI Receptionist
The front desk: greet, route, schedule.
Tools and trail
Clinics, practices and offices where the front line is one person who also has another job.
- Phone
- Calendar
- CRM
- Greet
- Route intent
- Book or transfer
- Log outcome
Anything medical, legal, or a complaint goes straight to a named person. It does not try to judge what it is not qualified to judge.
Sales / SDR Agent
Research → outreach → book → CRM.
Tools and trail
Reps who spend the first hour of the day researching accounts instead of speaking to any of them.
- CRM
- Web
- Target loaded
- Brief + draft
- Sequence send
- Book reply
Stops before the first outbound touch on any account a rep already owns.
Support Agent
Triage, resolve or draft, escalate safely.
Tools and trail
Tier-one queues that never shrink because the same twenty questions arrive again every morning.
- Helpdesk
- KB
- CRM
- Classify ticket
- KB match
- Resolve or draft
- Escalate low confidence
Escalates on a second failed resolution, on any refund or account deletion, and on detected frustration.
SMS / Text Agent
Confirms, chase, reactivation on SMS and WhatsApp.
Tools and trail
Teams who send a quote or a booking and then hear nothing, with no idea which silence is worth chasing.
- SMS
- CRM
- Trigger fired
- Text sent
- Reply classified
- CRM updated
Stops the sequence on any reply it cannot classify, and on the first sign the recipient wants out.
Research Agent
Trusted briefs from sources you allow.
Tools and trail
Anyone walking into calls cold because the brief takes longer to write than the meeting takes to hold.
- Web
- CRM
- Docs
- Sources allow-listed
- Brief + citations
- Push to CRM
- Human reviews
Flags the gap instead of guessing when a claim has no source it was allowed to read.
Ops / Workflow Agent
Checklists, sync, dunning, routing in the stack.
Tools and trail
Back offices where the first hour of the morning goes on retyping the same facts between tools.
- Sheets
- CRM
- Slack
- Checklist open
- Sync systems
- Exceptions only
- Human approves writes
Halts the run when a record fails to match on both sides, because a half-synced row costs more than a stalled one.
Booking Agent
Availability, book, remind, chase no-shows.
Tools and trail
Calendars filling with appointments that nobody confirms and a proportion that nobody turns up to.
- Calendar
- Phone
- SMS
- Request slot
- Live availability
- Confirm + remind
- Chase no-show
Hands off before it would double-book, and on any request to move an appointment someone else owns.
Email Agent
Sort, draft, label, and route the inbox, inside written limits.
Tools and trail
Founders whose inbox has become the queue everyone else in the company depends on.
- CRM
- Calendar
- Inbox classified
- Drafts prepared
- Labels + CRM
- Human before send
Drafts and stops on anything addressed to a customer or a regulator until you have read it.
Coding Agent
Plan, implement, test, and open PRs inside your repo rules.
Tools and trail
Teams with a backlog of well-specified work and acceptance tests already written against it.
- Git
- CI
- Issues
- Issue defined
- Branch + implement
- Tests run
- PR for human review
Opens a pull request and stops. It does not merge, and it does not touch migrations or secrets.
Content Agent
Brief, draft, SEO checks, then a publish queue, all in your brand voice.
Tools and trail
Distribution that stalls whenever the one person who can write is busy with something else.
- CMS
- Docs
- Analytics
- Brief locked
- Draft on voice
- SEO pass
- Human approves publish
Publishes nothing. The queue ends at a draft with a human name against it.
Data Analyst Agent
Queries, charts, and explains, with sources you can open and check.
Tools and trail
Meetings that wait on somebody pulling the numbers, then argue about which pull was correct.
- Warehouse
- BI
- Sheets
- Question parsed
- Query run
- Chart + note
- Human if anomaly
Reports the number as unavailable instead of approximating it when a source is stale or a join is ambiguous.
Browser / Computer-Use Agent
Navigates interfaces, fills forms, and extracts data where no API exists.
Tools and trail
Legacy portals and supplier systems with no API, where the work is a person clicking through screens.
- Browser
- Vault
- Logs
- Open target UI
- Complete steps
- Capture proof
- Stop on CAPTCHA / edge
Stops at any payment, any consent screen, and any CAPTCHA. It does not learn its way around a block.
Knowledge / RAG Agent
Answers from your own documents with citations attached, and says so when the answer is not in them.
Tools and trail
Teams sitting on a knowledge base that nobody searches because searching it has never once worked.
- Vector DB
- Docs
- Chat
- Retrieve chunks
- Answer + cite
- Gap flagged
- Human if no source
Answers that it is not in the documents instead of filling the gap from general knowledge.
HR / People Agent
Policy Q&A, onboarding checklists, leave routing.
Tools and trail
People operations answering the same policy question in writing several times a week.
- HRIS
- Docs
- Slack
- Request typed
- Policy cited
- Ticket or approve path
- Human for disputes
Routes anything touching pay, performance, or a grievance to a person, without recording an opinion of its own.
Finance Agent
Invoice matching, dunning drafts, and expense flags, all behind approval gates.
Tools and trail
Finance teams where the exceptions have quietly become the job and the routine work fits around them.
- ERP
- Bank
- Match documents
- Flag mismatches
- Draft chase
- Human before pay
No payment leaves without human approval, and any mismatch over the threshold you set stops the batch.
Legal Intake Agent
Classify requests, pull clauses, draft for counsel review.
Tools and trail
In-house legal buried in intake, spending its expensive hours sorting requests before it can answer any of them.
- Docs
- Matter
- Intake classified
- Sources attached
- Draft for counsel
- Never auto-send advice
Drafts for counsel. It gives no advice and it signs nothing.
Onboarding Agent
Customer or employee onboarding checklists end-to-end.
Tools and trail
Customer success and HR teams chasing setups that stalled two weeks ago without anyone noticing.
- CRM
- Docs
- Slack
- Checklist opened
- Tasks chased
- Blockers escalated
- Marked complete
Escalates on the second missed step, before the customer notices the setup has stalled.
Monitoring Agent
Watch signals, alert, and open the right ticket.
Tools and trail
Operations teams who find out about failures from a customer first and from their own monitoring second.
- Metrics
- Pager
- Slack
- Signal detected
- Severity scored
- Ticket + context
- Human on SEV1
Pages a person when an alert repeats inside its own cooldown, because a loop it cannot break is one it cannot see.
Lead Qualification Agent
Score and route inbound leads before sales touches them.
Tools and trail
Sales teams spending their week on inbound that was never going to fit in the first place.
- Forms
- CRM
- Chat
- Lead captured
- Score vs ICP
- Route or nurture
- Human for edge ICP
Routes a lead to sales unscored instead of discarding one it could not read.
Multi-Agent System
Orchestrator + specialists when one agent is not enough.
Tools and trail
Jobs complex enough that one agent doing everything in sequence would take longer than a person.
- Orchestrator
- Workers
- Logs
- Job decomposed
- Specialists run
- Results merged
- Human on conflict
The lead agent holds the hand-off to a person. Specialists cannot go past each other or around it.
Meeting Agent
Captures the call and puts notes, owners, and follow-ups back into the stack.
Tools and trail
Teams whose decisions are real but whose actions die somewhere inside an unread transcript.
- Transcript
- CRM
- Tasks
- Meeting recorded
- Actions extracted
- Owners assigned
- Human on disputes
Leaves an action unassigned instead of guessing an owner from a transcript.
Outbound Voice Agent
Reminders, renewals, and campaign dials with CRM write-back.
Tools and trail
Anyone working a manual dial list for reminders and renewals that follow an entirely predictable script.
- Phone
- CRM
- Scripts
- List loaded
- Call placed
- Outcome logged
- Human on opt-out / anger
Ends the call and suppresses the number on any request not to be contacted again.
Customer Success Agent
Health signals, renewals, and QBR prep for account teams.
Tools and trail
Account teams doing status archaeology before every review instead of arriving with the picture already assembled.
- CRM
- Product
- Health scored
- Risk flagged
- Playbook started
- Human on churn risk
Escalates a health drop to the account owner before it drafts anything the customer would see.
Recruiting Agent
Sources, screens, and schedules, with a human on every offer.
Tools and trail
Lean teams hiring steadily without a talent acquisition function to run the top of the funnel.
- ATS
- Calendar
- Role briefed
- Candidates screened
- Interviews booked
- Human on offer
Screens and schedules only. Offers, rejections, and anything a candidate disputes go to a person.
DevOps Agent
Runbooks, sorting incidents by urgency, and safe release help.
Tools and trail
On-call rotations where the same incidents recur and the runbook for them is already written.
- Pager
- CI
- Cloud
- Alert received
- Runbook matched
- Safe actions run
- Human on SEV1
Runs read-only against production until a named engineer confirms the step.
Code Review Agent
PR review for bugs, style, and security flags.
Tools and trail
Teams whose pull request queue is the bottleneck and whose reviewers are also the people shipping.
- Git
- CI
- Security
- PR opened
- Diff reviewed
- Comments posted
- Human on merge
Flags and blocks. It cannot approve away its own findings.
QA / Test Agent
Generate and run tests; file failures with context.
Tools and trail
Product teams shipping faster than a test suite can be extended by hand to keep up.
- Suite
- CI
- Tickets
- Change detected
- Tests planned
- Suite run
- Human on flakes
Files the failure with a reproduction. Whether a failing test ships is not its call.
SEO / GEO Agent
Briefs, clusters, and citation-ready pages for discovery.
Tools and trail
Companies with a real offer that search results and answer engines currently describe as nothing in particular.
- Search
- CMS
- Analytics
- Topic mapped
- Brief drafted
- Page shipped
- Human on claims
Stops at the brief when a claim on the page would need a measurement nobody has taken.
Compliance Agent
Policy checks and evidence packs before humans sign.
Tools and trail
Teams that rebuild the same evidence pack from scratch in the fortnight before every audit.
- Policy
- Docs
- Tickets
- Control selected
- Evidence gathered
- Gaps listed
- Human on attest
Assembles the evidence and stops. Every sign-off stays with a person who can be held to it.
Critique / Debate Pair
A worker and a reviewer agent, so the answer improves before anyone is asked to look at it.
Tools and trail
High-stakes drafts where being wrong is expensive enough to justify a second pass before a person reads it.
- Worker
- Critic
- Logs
- Draft produced
- Critic scores
- Revise or pass
- Human on dissent
When worker and critic disagree twice, both outputs go to a person with the disagreement attached.
After the board
Not on the board? Still possible.
Every engagement starts from the work you actually run. If the work repeats, if it can be done through tools, and if there is a point at which a human should take over, then it can be built as an agent and the board has nothing to do with it.
FAQ
Straight answers
Do you only build the agents on the Switchboard?
No. The board is a sample of the kinds people ask for, and a custom line is the ordinary case. If your job repeats, runs through tools, and has a point where a person should take over, it can be built whether or not it appears above.
How is this different from a chatbot?
A chatbot answers a question and the conversation ends there. An agent picks up a job, works through your systems to finish it, records what it did, and stops to ask for a person when it hits the condition you set. The difference shows up on the day something goes wrong: with a chatbot there is nothing to read.
Can you build multi-agent systems?
Yes, when the work genuinely needs parallel roles. One agent holds the job and specialists take pieces of it, with every hand-off written down and one accountable design behind the whole thing. Most jobs do not need this, and adding a second agent to a system that was not working with one usually makes the failure harder to find.
How long does a first agent take?
Typically 4 to 8 weeks for a focused production agent, after a short discovery that establishes what the job actually is. What gets built is written down and fixed before anything starts, so the timeline is a commitment and not an estimate that drifts.
What happens when the agent is wrong?
It stops on a condition agreed in advance and hands the job to a named person with the context attached, which is why every line on the board carries a different escalation rule. The trail then tells you what it did before it stopped, so the fix goes into the rules, where you can see it, and stays out of a retraining exercise nobody can audit.
Do our customers know they are talking to an agent?
Yes, on any voice or chat line I build. Faking a human is a refusal here, not a configuration option, and it is also the fastest way to lose the trust that makes the rest of the system worth having.
Where do you work?
The Netherlands and the wider EU, remote-first, with engagements running in English. The next start is typically within two weeks of a signed agreement.
Further reading
AI agents in the Netherlands
Dutch and EU teams, English engagements, remote-first. The Netherlands page covers governing law, what an engagement is contracted on, and where your data actually goes. Next start typically within two weeks.
Last reviewed