The work your team shouldn't be doing.

Fettler is an AI-services studio. We build custom agents, skills, automations and integrations for mid-market businesses, then run and monitor them inside your systems, under your approval.

Skilled people spend their day re-typing between systems that do not speak to each other. We build agents that do that work instead.

Working with 40+ clients across sizes and industries.

shift production and downtime capture, drawn end to endshift production and downtime capturestarts on shift closeshift logsheet and photoline planoutput targetsdowntimereason codesshipment plandue datesshift production agentunattended01Readshift entries02Decideplan and downtime03Actrecord and updatecodedno reason codeshift recordready to postshipment planupdate draftedApproval gateplant manager signs off the shiftsent and postedafter a person released itrun logevery input, rule and output shift production and downtime capture, drawn end to endshift production and downtime capturestarts on shift closeshift logsheet and photoline planoutput targetsdowntimereason codesshipment plandue datesshift production agent01Readshift entries02Decideplan and downtime03Actrecord and updatecodedno reason codeshift recordready to postshipment planupdate draftedApproval gateplant manager signs off the shiftsent and postedafter releaserun logevery input, ruleand outputshift production and downtime capture, drawn end to end Trigger sourcestarts on shift closeshift logsheet and photo Readshift entries Decideplan and downtimecoded no reason code Actrecord and update Approval gateplant manager signs off theshift sent and posted / run log

Production: shift production and downtime capture

  1. The shift ends and the log is submitted as a sheet, photo or chat message.
  2. Read output and downtime entries against the line and shipment plans.
  3. Compare output to the plan and check downtime reason codes. Downtime without a reason code is held for the shift supervisor to name.
  4. Prepare the shift record and update the shipment plan.
  5. The plant manager signs off the shift record before it is sent and posted.
  6. Log every input, rule, decision and output.

The trigger changes and the approver changes. The anatomy does not.

  • a system, a file or a message
  • work the agent does on its own
  • a decision against your rules
  • a named person has to release it

Demonstration data. Not a client system.

The problem

Most mid-sized businesses do not have a technology problem. They have too much work going through too few hands.

  • The hours don't go into making things. They go into re-typing things.

  • Buying another module moves the typing, it does not remove it.

  • The exceptions never make it into any system.

How this differs from the usual programme.

  • Scope

    The usual way
    A platform rollout that touches every department at once.
    With us
    One workflow at a time, each paying for itself before the next.
  • Time to real work

    The usual way
    A programme measured in quarters, with a pilot at the end.
    With us
    The first agent doing real work by week six.
  • Proof

    The usual way
    A demo on sample data and a steering committee sign-off.
    With us
    Parity with your manual process, reconciled to zero difference, before anything is switched off.
  • Pricing

    The usual way
    Seats, licences and time-and-materials meters.
    With us
    A fixed price per workflow and a monthly fee to run it.

What we deliver

Custom AI agents, skills, automations and integrations.

You buy one workflow at a time, not a platform to grow into. Our agentic AI services cover the workflow, the connections and the work of keeping it running.

Agents
A digital worker that owns a whole workflow end to end.
builds the full export documentation pack and stops on any mismatch
Skills
One reusable competence that any agent or any person can call.
one "read a supplier invoice" skill, every vendor, every layout
Automations
Scheduled unattended runs, with monitoring and retries.
the daily close pack is on the desk before the desk is occupied
Integrations
Getting data out of systems never meant to share it: a legacy ERP, a supplier portal, bank mail, PDFs.
platform integration where no API was ever published

Custom AI agent development

The machinery

Every AI agent we build has the same six parts.

Name the trigger, the rules and the person who signs.

Anatomy of an agent: trigger, read, decide, act, escalate, log 01 02 03 04 05 06 Trigger Read Decide Act Escalate Log it starts itself ERP, PDFs,portals, mail,sheets, photos your rules, yourtolerances, yourdefinitions posts, files,drafts, emails,updates ambiguity goesto a person,never a guess every input,rule and output,auditable Anatomy of an agent: trigger, read, decide, act, escalate, log 010203 040506 TriggerReadDecide ActEscalateLog it starts itself ERP, PDFs, portals,mail, sheets, photos your rules,your tolerances,your definitions posts, files, drafts,emails, updates ambiguity goes to aperson, never a guess every input, rule andoutput, auditable
  1. Trigger: it starts itself
  2. Read: ERP, PDFs, portals, mail, sheets, photos
  3. Decide: your rules, your tolerances, your definitions
  4. Act: posts, files, drafts, emails, updates
  5. Escalate: ambiguity goes to a person, never a guess
  6. Log: every input, rule and output, auditable

Not a chatbot

A chatbot answers. An agent finishes the task: the document is filed, the entry is posted.

Not recorded clicks

A recorded click sequence breaks the week a screen moves. An agent reads meaning.

Not a dashboard

A dashboard waits to be opened. An agent acts on the number and tells the person who needs to know.

Already running

AI workflows already running in business operations.

We describe them by what they do rather than by who they belong to.

  • Finance

    Settlement reconciliation

    Reads dense remittance workbooks by meaning rather than cell position and rebuilds every line; anything it cannot account for is held with the reason stated.

    trigger
    remittance file lands in the mailbox
    approver
    finance lead
  • Leadership

    An AI colleague on chat

    Answers plain-language questions against the live database using one recorded formula per metric, in the chat leadership already uses.

    trigger
    a question is asked
    approver
    none, read only
  • Operations

    The watchdog

    Sweeps every scheduled process and alerts the owner, not the room, with the fix in the message. Silence means healthy.

    trigger
    continuous
    approver
    process owner
  • Finance

    Bank to invoice to filing

    Matches credit alerts to open invoices, names every deduction, raises the tax invoice and produces a return-ready workbook.

    trigger
    bank credit alert
    approver
    finance controller
  • Customer service

    Inbound at scale

    Classifies and routes thousands of messages a week to an owner, escalates critical language within minutes, drafts replies a person releases.

    trigger
    message received
    approver
    named responder
  • Integration

    The collection layer

    Scheduled jobs that sign in to portals with no published interface, with session management, retries and a per-run audit trail.

    trigger
    schedule
    approver
    none, read only

Scroll sideways to read the table.

Demonstration data. Not a client system.
timeagentoutcomestate
06:00settlement recon4 remittance files read, every line rebuiltlogged
06:14settlement reconone deduction unnamed, held for the finance leadescalated
07:30supplier invoice61 invoices matched to purchase orderslogged
08:00compliance sweep12 checks due today, evidence attachedlogged
09:12payment postingcredit matched to invoice, tax invoice draftedawaiting approval
09:40inbound triagemessages classified, 3 critical routed to ownerslogged

Control

The agent prepares. A person releases.

Anything that leaves your company or moves your money has a named approver.

Scroll sideways to read the table.

Demonstration data. Not a client system.
raisedactionapproverstate
09:12release tax invoice to customerfinance leadheld for approval
09:31post credit note, ledger writefinance controllerheld for approval
10:58send documentation pack to buyerexport deskreleased 11:04
nothing in this queue moves without a named human
The agent prepares, a named person releases, and only then does the work leave your businessSix guardrails hang off the gate: data, authority, evidence, integrity, continuity, reversibility.The agentprepares the workThe gatea named personreleases the workYour businessand everyone outsideDataAuthorityEvidenceIntegrityContinuityReversibility The agent prepares, a named person releases, and only then does the work leave your businessSix guardrails hang off the gate: data, authority, evidence, integrity, continuity, reversibility.The agentprepares the workThe gatea named personreleases the workYour businessand everyone outside
Data
Your cloud, your network, or your own hardware.
Authority
A human holds the pen.
Evidence
What it read, which rule it applied, what it produced, and who approved it.
Integrity
No fallback logic. If a field is missing it stops and asks.
Continuity
Documented rules and integrations, exportable data, and a handover pack.
Reversibility
Parity is proven before anything is switched off.

We enter your systems only with your written authorisation and your own credentials.

Engagement

Four phases. The first agent is live on real work in six weeks.

Four phases across nine weeks, with the parity check drawn as a fork that rejoins01Diagnostic, 2 weeks02Build, 4 weeks03Prove, 3 weeks04Run and extend, ongoingweek 0269ongoingthe agent runsthe manual process runs beside it until the two agree Four phases across nine weeks, with the parity check drawn as a fork that rejoinsweek 0269ongoingagent and manual, side by side
The four phases of an engagement, with how long each takes
#PhaseWhat happensWhat you have at the end
01 Diagnostic, 2 weeks We sit beside the people doing the work and map eight to twelve workflows, measuring what each one costs in hours. A ranked automation backlog with an expected payback against every workflow.
02 Build, 4 weeks The first agent is built against the real process, in your environment, on your data. Weekly demonstrations with the people who will use it, not with a steering committee. One agent in production, doing real work.
03 Prove, 3 weeks The agent runs alongside the manual process and the two are reconciled to zero difference. Evidence, and then the switchover.
04 Run and extend, ongoing We host it, monitor it, maintain it, and absorb the format and portal changes that break automation. The next workflow comes off the backlog every four to six weeks. Compounding coverage rather than a finished project.
  • Numbers are set after the diagnostic, because quoting before we have seen your process goes badly.

How the diagnostic and build work

Fit

If your work is rules based, repeated and spread across systems that do not talk, we can automate it.

  • Skilled people spend more than half their week moving information between systems.
  • The number everyone actually trusts lives in one person's spreadsheet.
  • The deadline is met by someone staying late.
  • The last automation attempt produced a pilot, and the pilot produced a meeting.
  • Manufacturing
  • Shipping and logistics
  • Hospitality
  • Multi site consumer operations
  • Business services

What that looks like in a manufacturing business

A first diagnostic in a contract manufacturing business produced this backlog in two weeks.

  • Export documentation packs
  • Purchase to pay matching
  • Customer purchase order to sales order
  • Production against shipment plan
  • Shift production and downtime capture
  • Inventory replenishment triggers
  • Inline quality checks with photo evidence
  • Complaint to corrective action
  • Tool and die shot count maintenance
  • Request for quotation to quotation
  • Payables and receivables follow up
  • The compliance calendar

We would expect at least three of these to be wrong.

AI workflow automation for manufacturing

Method

Four rules we do not break, and the reason each one exists.

MIT's Project NANDA reported in 2025, in The GenAI Divide, that about 95 per cent of the organisations it studied saw no measurable return from generative AI. We read that as a discipline problem more than a technology problem. These are the four disciplines we hold ourselves to.

  • No fallback logic

    An agent never quietly substitutes one data source for another when a field is missing.

  • A percentage is not an explanation

    A difference is closed when we can say what every unit of it is, not when the percentage looks small. Anything left over is reported as unexplained.

  • Parity before switchover

    The agent runs beside the manual process until the two reconcile to zero difference. If it does not tie out, it does not go live.

  • The done-ness test

    After the agent goes live, does a person still open that file? If yes, we have not finished.

Pick one workflow. We'll have it running in six weeks.

Two weeks beside your team, eight to twelve workflows mapped and costed, and a ranked backlog with an expected payback against each one.

Book the diagnostic

Or write to us and describe the workflow that annoys you most. That is usually the right one.

hello@fettler.ai

Questions

Eight questions, answered the way we would answer them on a call.

Where does our data live?

Wherever you tell us: your cloud account, your network, or your own hardware. We do not pool it with anyone else's and we do not use it to train a model.

What do you get access to?

The least we need for the workflow you asked for, granted in writing, using credentials you issue and can revoke. We tell you what an agent touches before it touches it.

Who owns the code?

You do, for everything built specifically for you: the workflow logic, the rules, the integrations and the documentation. This is written down before the first build, not after.

What happens if we stop working with you?

You get a handover pack: documented rules and integrations, credentials returned, data exported in a format you can read, and a walkthrough. We build for that day from the beginning.

How do you price this?

A fixed fee for the diagnostic, credited against the first build. A fixed price per agent and a monthly fee to run it. We do not sell seats or bill by the hour.

What if it gets something wrong?

If it is uncertain it does not proceed: it stops and escalates to the person you named. If it acts wrongly, the log tells you what it read and which rule it applied.

Do our people have to learn a new tool?

No, and we treat that as a design constraint. The work shows up where your people already work: the ERP screen they use, their mailbox, the chat they already have open.

How long before something is live?

Two weeks for the diagnostic, and the first agent is on real work by week six. Parity is proven between weeks six and nine before anything manual is switched off.