Sidreal

The consequence layer for AI agents

Read the reply.
Forecast what it causes.
Then send.

Sidreal reads each reply your AI agent drafts, checks what it commits you to against your own policy, and forecasts whether it ends in a reopen, an escalation or a complaint.

Selecting design partners · Frankfurt am Main

Watch one reply

One reply, rewound

Your agent answers a customer after a four-day outage, politely and with every fact right. Four of its five sentences come back to cost you. Sidreal reads them before the reply goes out.

Your rules

Goodwill limit
Credits post within

Drafted by your agentnot sent · forecast

Subject
Your outage from 3 to 7 March
Hello Ms Keller, I am sorry for the outage from 3 to 7 March. I have credited €60 to your account. Under our terms, this outage does not qualify for a free cancellation. The credit will appear within 3 working days. For further questions, please see our FAQ.

  1. Sent Commitment €60 credit promised. Your goodwill limit is €25.
  2. +40 min Complaint “Four days without internet, and I can’t even cancel?”
  3. +4 days Reopen The credit is not on the account. Same customer, same ticket.
  4. +6 days Escalation The customer asks for a team lead.

What the person sees

    No marks. It can go out as written.

    Draft Sent +40 min +4 days +6 days
    Illustration. The reply, the rules and everything after them are written for this page. Press Send, change a rule, or drag the mark along the time line.

    What your stack already answers

    Your content filter and access policy would let this reply through. Most checks ask whether a reply is allowed. Some score a case once it is closed. Sidreal reads what a single reply will cause, before it goes out.

    Sidreal checks commitments against the policy you supply. It does not verify other facts, legal questions go to a person, and medical content stays out.

    Per draft, before it is sent Per case, after sending Is it allowed?rules, intent, policy, sign-off What will it cause?reopen, escalation, complaint, satisfaction

    Per draft, before it is sent · Is it allowed?

    • Platform supervisorsThe same vendor writes the reply and checks it.
    • Approval inboxesA person approves each action.
    • Policy-rule verifiersText checked against written rules.
    • Full human reviewCosts the time the AI was meant to save.
    • Sidreal commitment checkWhat a reply promises, against your rules.

    Per case, after sending · Is it allowed?

    • Sampled QATeam leads score closed tickets. The customer already has the reply.

    Per draft, before it is sent · What will it cause?

    • Sidreal outcome forecastWhat one draft will cause, and how sure it is. Outside the platform that wrote it.
    • Your own judge promptWritten in an afternoon. Sidreal is measured against it.

    Per case, after sending · What will it cause?

    • Escalation predictionPer case, from sentiment and history.
    • Satisfaction scoresPer conversation, after it ends.
    Categories from vendor documentation and press, October 2026.

    It already happens

    One reply can cost more than the agent saves.

    Feb 2024Commitment

    Air Canada

    A tribunal held Air Canada to what its website chatbot had told a grieving customer about bereavement fares. The airline had argued that the chatbot was responsible for its own actions.

    What a reply promises, the company owes.

    Moffatt v. Air Canada, 2024 BCCRT 149

    Apr 2025Cancellations

    Cursor

    The support bot explained a login problem as a one-device policy. No such policy existed. Users threatened to cancel on Reddit and Hacker News before a co-founder wrote: “We have no such policy.”

    Sidreal would not have caught the invented rule. A reply that announces a restriction to paying customers is what the outcome forecast reads.

    Report in CX Today

    May 2025Switched back

    Klarna

    After its AI assistant took over most customer chats, Klarna’s CEO said customers should always be able to reach a person, and the company started hiring agents again.

    The pullback Gartner predicts, at the company that went furthest.

    Report in Entrepreneur

    May 2026Forecast

    Gartner

    Gartner expects 40% of enterprises to demote or decommission autonomous AI agents by 2027, after governance gaps surface in production incidents.

    Its fix sorts agents by autonomy. Sidreal sorts the replies one agent writes.

    Press release

    1. Commitment check

      What a reply promises, against the rules you sign off.

      1. Designed
      2. Prototype
      3. Pre-tests
      4. Real replies
      5. Field-proven

      Now Pre-tests on constructed replies

    2. Outcome forecast

      Risk of a reopen, escalation or complaint, fitted on your own history.

      1. Designed
      2. Evaluation
      3. Backtest
      4. Shadow run
      5. Field-proven

      Now Run end to end on public support replies. Next: your history

    3. Replayable runs

      Every check is stored with its input, model version and output, so it can be replayed later.

      1. Designed
      2. Built
      3. In use

      Done In use for every result we show

    4. EU-hosted reading model

      The model that reads the draft runs in the EU.

      1. Hosted in Hesse
      2. Contracted provider
      3. Field-proven

      Now Provider under contract for paying customers

    5. Runs in your network

      The whole check inside your network, with no outbound calls.

      1. Planned
      2. Local alpha
      3. Package
      4. Licence

      Now Local alpha

    Updated with every release.

    Built by

    Gregor Pestsov

    Gregor Pestsov Founder

    Twelve years validating systems before they go live, in MedTech, automotive and food tech: FDA compliance and computer systems validation. MBA in Digital Business, with a thesis on AI in enterprise risk management.

    In regulated industries, nobody signs off the release of a system they built themselves. Sidreal brings that rule to AI agents.

    LinkedIn profile

    Next step

    Where is your AI agent today?

    About how many AI replies a month?

    Which outcomes can you tie to a single reply?

    Where a first call would start

    Your message to Gregornot sent

    To
    gregor@sidreal.ai
    Subject

    Add your company and anything about your setup. Nothing leaves this page until you send it.

    Send from your mail app or Outlook Gmail

    Or write directly: gregor@sidreal.ai