M18
Replay my tickets

Support, resolved.

Autonomous support for your machines. A model trained on every fault you ever fixed, on your servers. Your engineers only see what is new.

Replay 1,000 of your tickets See the benchmark
ServiceNow · INC-20931solved

R. Moreau · customer portal · 6:12 am

“Charger B-12 shows E-47, no session starts”

  1. Asset identifiedB-12 · DC150-0847 · under contract
  2. Logs read, cause foundE-47 · known path, 41 cases
  3. Procedure checkedP-118 · remote restart allowed
  4. Rehearsed, then executedtest unit ✓ · controller restarted
  5. Back in service, ticket closed150 kW delivered · solved
96 s · 0 technicians · 1 resolution billed
  • ≥ 70 %of tickets resolved end to end: the launch bar
  • < 60 sfirst response, 95th percentile, 24/7
  • 100 %of answers traced to a source in your own data
  • On-premyour model and its memory run on your servers

M18 commitments, written into the contract and checked on a replay of 1,000 of your closed tickets. Below the bar, you owe nothing.

The model and its memory

A model trained on your data. A memory made of your know-how.

Your own model, trained on your tickets, manuals and procedures, running on your servers. Its memory is the graph below. The model reasons. The memory proves.

  • 18,412nodes
  • 61,307links
  • 9,880sources
  1. 01

    Your model

    Trained on your closed tickets, manuals and procedures. It speaks your machines, your codes, your vocabulary. It belongs to you.

  2. 02

    Its memory

    A knowledge graph of every fault you ever solved. Every fact keeps the document it came from.

  3. 03

    No path, no answer

    Every answer is a path through the graph, from symptom to cause to fix, with its sources. Where no path exists, the agent does not guess. It hands over.

  4. 04

    It compounds

    Every closed ticket trains the model and grows the memory. When an expert retires, what they knew stays.

Illustration. Your model and your graph are built from your own data during the replay.

Autonomy

Five levels of support autonomy.

Defined by who closes the ticket and what the system may touch. The market ships L1 and L2.

  1. L0

    FAQ bot

    Canned answers. Deflects, never resolves.

    Help centres
  2. L1

    Copilot

    Drafts the reply. An engineer still sends it.

    Helpdesk add-ons
  3. L2

    Resolver

    Answers from your manuals. Escalates every action.

    Generalist AI agents
  4. L3

    Operator

    Acts within your procedures. Hands over the rest.

    The best of 2026
  5. M18 operates hereL4

    Autonomous operation

    Resolves, acts, prevents the next fault. Every answer sourced in your memory.

    M18

Level definitions and the tests behind them are part of the benchmark protocol, shared in full before the replay.

Three layers

Before the ticket. Inside the ticket. After the ticket.

A support operation is not one agent answering emails. It is three loops running on the same memory of every fault you ever closed.

  1. Layer 0 · Before

    Resolved while the operator types

    On your customer portal, in your app, next to the machine: the operator describes the fault, the fix appears before a ticket exists. Most routine faults end here.

  2. Layer 1 · Inside

    Resolved and acted, end to end

    Asset identified, logs read, procedure applied, remote action rehearsed then executed, ticket closed. Or a technician dispatched with the right part and the diagnosis.

  3. Layer 2 · After

    The next fault, prevented

    Every week, the recurring faults ranked by tickets avoided, sent to engineering and field service with the fix. The volume falls month after month.

No more L1. No more L2.

What a technical support team gains. Everything you already measure, moving the right way.

ServiceNow · INC-20931solved

R. Moreau · customer portal · 6:12 am

“Charger B-12 shows E-47, no session starts”

  1. Asset identifiedB-12 · DC150-0847 · under contract
  2. Logs read, cause foundE-47 · known path, 41 cases
  3. Procedure checkedP-118 · remote restart allowed
  4. Rehearsed, then executedtest unit ✓ · controller restarted
  5. Back in service, ticket closed150 kW delivered · solved
96 s · 0 technicians · 1 resolution billed
L1 and L2 resolved end to endThe asset identified, the logs read, the procedure applied, the ticket closed.
tonight · first responseSAMPLE
  1. Charger offline after a power cut21:41 → 21:42
  2. E-47, no session starts23:07 → 23:08
  3. Card reader not responding01:30 → 01:31
  4. Session stops at 80 %04:15 → 04:15
Answered in seconds, 24/7At 3 am on a Sunday as on a Monday morning. No on-call rota for routine faults.
queue · liveSAMPLE
waiting0SLA 100 % · 7 days
Zero backlogEvery ticket picked up the minute it arrives. SLAs met without a night shift.
tickets per day · update 3.2 roll-outSAMPLE
peak×50 hires
Peaks without hiringA faulty update, a cold snap, a new site going live: five times the volume, same team.
Your team on L3Unseen faults and critical sites reach an engineer with the logs, the history and the attempts already attached.
same reply · 4 languagesSAMPLE
  • Restart the controller: Settings › Service › Restart.
  • Redémarrez le contrôleur : Réglages › Service › Redémarrer.
  • Starten Sie die Steuerung neu: Einstellungen › Service › Neustart.
  • Reinicie el controlador: Ajustes › Servicio › Reiniciar.
Every languageThe operator writes in theirs. The agent answers in it, with your procedures.
QA · every conversationSAMPLE
  • Procedure98
  • Accuracy99
  • Safety rules100
  • Resolution96
QA on 100 % of conversationsTone, accuracy, policy: every reply scored, not a sample.
top faults · this weekSAMPLE
  1. E-47 after update 3.2412
  2. Offline after a power cut286
  3. Card reader not responding131
  4. Session stops early97

→ sent to engineering

Weak spots surfaceThe recurring faults, counted weekly, handed to engineering and field service with the fix.

Technology

Every action rehearsed. Every answer sourced. Every step on the ledger.

Six components, one loop, around one memory. Runs on your own servers with zero egress, or in M18’s EU or US cloud.

  1. 01

    Channels

    Customer portal, email, app, the machines’ own alerts. One queue.

  2. 02

    Memory · knowledge graph

    Built from your tickets, manuals, procedures and logs: machine, symptom, cause, fix, part, expert, linked. Every fact sourced.

  3. 03

    Your model

    Trained on your data, on your servers. Finds the path through the memory, cites every node, scores its confidence. No path, no answer.

  4. 04

    Sandbox

    Replays the planned actions on a test unit or a copy of your systems. What fails is never proposed.

  5. 05

    Policy gate

    Hard limits per action: remote command, parameter, part, site. Below confidence, it hands over with the investigation attached.

  6. 06

    Actions & ledger

    Executes in your ticketing, remote-access and parts tools. Every step signed, timestamped, replayable.

06 → 02Every closure trains the model and writes back to the memory. The next similar ticket starts from the answer, not from zero.

  • < 60 sfirst response, 95th percentile
  • 100 %of actions rehearsed before they run
  • 100 %of answers traced to a source
  • 0actions outside the limits you set
  • On-premyour servers, zero egress · or M18 cloud EU / US
  • 0training on your data for anyone else

M18 Support Bench · v1.0

An open benchmark. Run on your tickets.

Published rates are averages on someone else’s customers. Ours is a protocol: seven indicators, a written definition each, a launch bar and a target. It runs on 1,000 of your closed tickets before you sign.

IndicatorDefinitionLaunch barTarget
Resolved end to endClosed by the agent, action done, not reopened within 7 days≥ 70 %85 %
Correct diagnosisSame cause as your engineers’ closure, judged blind by two readers≥ 95 %99 %
First responseTime to the first substantive reply, 95th percentile≤ 60 s≤ 20 s
Actions in policyRemote commands, parameters and parts inside the limits you set100 %100 %
Wrong closureClosed by the agent, reopened by the customer within 7 days≤ 3 %≤ 1 %
Handoffs investigatedHanded over with asset, logs and attempts attached100 %100 %
Sourced answersEvery answer is a path in the memory, with the tickets and documents it comes from100 %100 %
Blind replay

Each ticket restarts from the customer’s first message. The agent never sees the cause, the reply or the closure.

Time cut-off

For every ticket, the memory holds only tickets closed before it was opened. Nothing leaks from the future.

Judge and runs

The judge is your team’s closure. Two human readers on 100 random tickets. Three runs, mean and spread published.

Resolution rate, week by week of the replaySAMPLE
Sample report. Yours comes from your own tickets.
Answers with a complete path in the memorySAMPLE
  1. 54 %M1
  2. 68 %M3
  3. 79 %M6
  4. 87 %M12
  5. 93 %M24
Sample. The memory compounds: every closure adds to it.

Protocol v1.0, November 2026. Definitions, scoring and the report format are shared with you in full before the replay.

Replay my tickets

The deal

Replay. Build. Run.

Cancel any month. Your tickets, your memory and your agent stay yours.

  1. Weeks 1–3

    Replay on 1,000 tickets

    Your closed tickets, our protocol, your report. Below the bar on any indicator, you owe nothing and keep the report.

  2. Weeks 4–6

    Build, fixed price

    Connected to your ticketing, your remote access, your asset register and your parts catalogue. Procedures and limits written with you. Live in 30 days.

  3. Every month

    Run, per resolution

    A ticket resolved end to end is billed. A ticket handed to your team is not. The seven indicators published to you every week.

Compliance

  • GDPR
  • DPA & NDA first
  • On-prem, EU or US
  • Full audit log
  • Human in the loop
  • Your data trains only your model
  • SOC 2 ready

Tools

Works inside the tools you already run.

No migration. The agent picks up tickets where they arrive, reads your logs and acts through your remote access, asset register and parts catalogue.

+ any tool with an API or a login

  • SnServiceNow
  • Salesforce
  • SAPSAP
  • D365Dynamics 365
  • IFSIFS
  • JiJira Service Mgmt
  • Zendesk
  • TeMicrosoft Teams
  • OuOutlook
  • IoTOPC UA · MQTT

Replay your last 1,000 tickets.

Three weeks. Seven numbers, in writing. Below the bar, you owe nothing.

Sample
1,000 of your closed tickets
Bar
≥ 70 % resolved end to end
Judge
Your team’s closure, blind
Duration
3 weeks
Start the replay