Managed Workflow Operations · Workflow Rescue

It worked in the demo. It does not work on Tuesdays.

Five days on one workflow that is failing in production. We reproduce the failures instead of taking your word for them, classify them against a failure taxonomy, fix what fits inside an agreed scope, and hand you regression tests, monitoring and a runbook. You keep everything, including the honest list of what we did not fix.

Scope and duration agreed before work starts Artifacts are yours regardless of outcome Continuing afterwards is a separate decision

The build is not usually where it breaks.

The build goes fine. The demo is convincing — it has to be, it was run against inputs chosen because they work. Then it meets a month of real traffic, and the shape of the failure is nearly always the same: it is not that the workflow is wrong, it is that nobody defined what happens when reality does not match the assumptions, and nobody was watching when it started to.

By the time somebody calls it broken, the usual state is a workflow that half-runs, a person quietly redoing its output by hand, and no agreement on whether the problem is the tool, the build or the process. Five days of evidence resolves that argument faster than another month of opinion.

Nine ways a working automation stops working.

You will probably recognise yours in the first three. Every finding in the report is classified against this list, with the input and the log line that proves it.

Untested input shapes
It was built against five clean examples. Production sends the sixth: a scanned fax, an empty field, a name with a comma, a 40-page attachment. This is the first thing we look for.
No exception path
Items that do not fit the rules are silently dropped or pile into a queue nobody owns. The workflow reports success while quietly losing work — expensive precisely because it is invisible.
Silent breakage
A step fails and nothing alerts. Discovery happens weeks later when a person notices a downstream gap. No monitoring means no failure — until there is.
Upstream drift
A vendor changed an API, a form gained a field, a template changed. The workflow was correct against last quarter’s reality.
Non-determinism treated as a bug
A model step produces a slightly different answer each run and the flow was designed as if it would not. The fix is usually structural — constrain the output, verify it, decide what happens on disagreement — not a better prompt.
No verifiable end state
Nobody defined what "done and correct" means, so nobody can tell whether a run succeeded. This one is fatal to measure around and has to be fixed first.
Credential and permission decay
A token expired, a service account lost access, a seat was deprovisioned when somebody left.
Volume shape change
It handles ten items and collapses at four hundred, or handles a steady trickle and collapses on the month-end spike.
Ownership vacuum
The technical cause is minor; the real cause is that the person who built it left and nobody inherited it. This one is not a technical finding at all.

What happens, day by day.

Short and fixed on purpose. A diagnostic that can expand is a diagnostic that will.

  1. 01

    Day 1 — Access and reproduction

    Read-only access to the workflow, its logs and a sample of real inputs. We reproduce the failures rather than taking a description of them. If a failure cannot be reproduced, that is itself a finding and it is reported as one.

  2. 02

    Day 2 — Classification

    Every observed failure is placed against the taxonomy above, with evidence: the input, the run, the log line. You get counts, not adjectives — which failure mode accounts for what share of the breakage.

  3. 03

    Day 3 — End state and test set

    We define what a correct run means for this workflow and build a test set from real inputs, including the ones that broke it. Without this you cannot tell a fix from a coincidence.

  4. 04

    Day 4 — Bounded fixes

    We fix what fits inside the agreed scope, highest-cost failure mode first. Anything larger is written up with an estimate rather than silently started — this is a diagnostic, not a rebuild in disguise.

  5. 05

    Day 5 — Handover

    Regression tests, a runbook naming what to check and who to call, monitoring and alerting on the steps that matter, and a written report of what was fixed, what was not, and what it would take.

Deliverables, including the uncomfortable one.

Failure report
Every failure observed, classified, with the input and log evidence attached. Yours to keep whatever happens next.
Test set
Real inputs including the ones that broke it, with expected end states. This is the asset that survives longest.
Fixes
What fitted inside the bounded scope, applied and verified against the test set.
Regression tests
So the next change tells you it broke something, instead of production telling you a fortnight later.
Monitoring
Alerting on the steps where a silent failure would be most expensive. Configured in the tools you already use.
Runbook
What to check, in what order, who to call, and what "normal" looks like. Written for whoever inherits it, not for us.
What we did not fix
Written down explicitly, with an estimate. A diagnostic that only reports good news is not a diagnostic.
Commercial shape
Scope and duration are agreed before work starts, and quoted in writing. Gridex does not publish its terms; you get them with the quote rather than discovering them afterwards. Whether anything continues after the diagnostic is a separate decision, made once you have seen the work.

Worth doing, and not worth doing.

This is for you if

  • A workflow that already exists and already runs — this is a rescue, not a build.
  • Failures that are real and recurring, not a one-off outage.
  • Somebody who can give read access to the workflow, its logs and real sample inputs.
  • A workflow whose output actually matters to somebody downstream.
  • Willingness to hear that the honest answer is a rebuild.

Do not buy this if

  • Nothing built yet. There is nothing to diagnose — start with a scoping conversation instead.
  • You want a full rebuild inside five days. Say so and we will scope that separately.
  • No access to logs, inputs or the workflow itself. We would be guessing.
  • The workflow is already switched off and nobody misses it. Let it go.
  • You need someone to agree it is the vendor’s fault. We report what the evidence says.

The questions buyers actually ask.

Five days is short. Is this real?

It is deliberately short because the alternative failure mode is a discovery engagement that bills for a month and produces a slide deck. Five days is enough to reproduce failures, classify them and fix the largest bounded ones. It is not enough to rebuild a system, and we do not pretend otherwise — anything larger comes back to you as an estimate, not as scope creep.

What if you cannot fix it?

You still keep the failure report, the test set and the written assessment of what a real fix would take. That is a genuinely useful outcome: many teams discover the honest answer is that the workflow should not exist in its current shape, and knowing that in five days is cheaper than finding out in six months.

Do you need production access?

Read access to the workflow, its logs and a representative sample of real inputs. Write access only where a fix requires it, scoped and agreed in advance. If sensitive data is involved we work against redacted or synthetic inputs that preserve the failing shape.

Does it matter what it was built with?

Rarely. The failure modes above are platform-independent — untested inputs, no exception path, no monitoring, no defined end state. Those look the same in Zapier, Make, n8n, a Python script, a vendor platform or a custom build.

What does it cost?

Quoted per workflow before anything starts, based on scope rather than hours. Gridex does not publish a price or its terms, because complexity varies enormously and because neither has been approved for publication. Describe the workflow and you get the number and the terms in the first reply.

What happens afterwards?

Whatever you want. Many teams take the runbook and tests and run the workflow themselves — that is a legitimate and good outcome. If you would rather somebody else owned running it, that is managed operations, and it is a separate decision made after you have seen how we work.

Have you done this before?

Not with a published case study. This offer is being validated — Gridex is deliberately selling a small number of these to find out whether the demand is real before building anything around it. The diagnostic itself is bounded, produces artifacts you keep, and is priced so a first engagement is a small decision.

Send us the workflow and what it does wrong. You get the number in the first reply.

Describe the workflow