Case Studies

Systems quietly running in the background

Every figure below is measured against a 90-day baseline taken during the audit, before anything was built. Where a number moved for reasons other than the system we shipped, we say so.

0Workflows shipped
0Hours saved / year
0Avg. ROI in year one
0To first automation live
Logistics · 40 staff

Cut quote turnaround from 3 days to 11 minutes

A freight broker was losing tenders to faster competitors. Not on price — on the three days it took to get a number back to the customer.

The problem

Inbound RFQs arrived as email, most of them unstructured and many as attachments. An ops coordinator re-keyed each one into a spreadsheet, checked rate cards across two systems, phoned a depot to confirm capacity, then wrote the quote by hand. Volume had tripled in two years; headcount had not.

What we built

  • An extraction agent that reads the inbound email and any attachment, and pulls out lane, weight, dates and special handling into a structured record.
  • A pricing step that queries the existing rate cards and the capacity system directly, rather than through a person.
  • A drafted quote surfaced in their CRM with the reasoning shown, for a human to approve or amend in one click.
  • A confidence threshold — anything unusual is escalated rather than guessed at, which is what kept the error rate down.

What we would do differently

We spent three weeks on attachment parsing before discovering that 80% of the awkward cases came from two customers who could simply be asked to use a form. Cheaper fix, found too late.

"We'd tried automation twice before and both attempts rotted within a month. Mindspace built the boring parts properly — logging, retries, alerts — so it's still running a year later."
Leif ElgethumFounder & CEO, Retrolux
B2B SaaS · Series A

Support deflection without the angry tickets

Deflection projects usually trade customer satisfaction for cost. The brief here was explicitly that CSAT must not move.

The problem

A three-person support team was handling 900 tickets a month, roughly two thirds of which were variations on the same dozen questions. Hiring was the obvious answer and the one the founders wanted to avoid — the queue was seasonal, and the questions were answerable from documentation that already existed.

What we built

  • A retrieval agent grounded in four years of resolved tickets plus the current help centre — not the marketing site, which was the source of most early wrong answers.
  • An evaluation set of 200 real historical tickets with known-good replies, run against every prompt change before it shipped.
  • A confidence gate that hands over to a human the moment the answer is uncertain, with the draft attached so the agent starts from something.
  • A weekly digest of deflected topics, which the team now uses to decide what documentation to write next.

Honest caveat

Deflection sat at 41% for the first six weeks. The jump to 62% came from rewriting eleven help articles the agent kept stumbling on — the content work mattered as much as the system.

"The audit alone was worth the engagement. They talked us out of two ideas we were excited about and pointed at one we'd completely overlooked. That one now saves us six figures."
Bryan PlasterEntrepreneur
E-commerce · 8-figure

One dashboard replacing nine spreadsheets

No AI in this one. The founder asked for a forecasting model and left with a data pipeline, which is what the problem actually called for.

The problem

Orders lived in the store platform, ad spend in three ad accounts, returns in a helpdesk, and margin in a spreadsheet maintained by one person. Every number the founder saw was at least four days stale, and two of the nine spreadsheets disagreed about revenue.

What we built

  • Nightly pipelines pulling orders, ad spend, returns and cost-of-goods into a single warehouse with one agreed definition of revenue.
  • Reconciliation checks that flag when two sources disagree, rather than silently picking one.
  • A morning digest in the founder's inbox: yesterday's numbers, week-to-date, and anything that moved more than a set threshold.
  • A dashboard for the wider team, built after the digest proved which numbers people actually opened.

Why no model

We scoped a demand forecast in the audit and recommended against it. With four-day-stale, internally contradictory inputs, a forecast would have been confidently wrong. That remains on the roadmap now that the inputs are trustworthy.

"Our team of four now handles the volume we were quoting eleven people for. No drama, no giant migration — they just shipped every other week."
Anil MathewInsurance Agent

Book a Audit

Curious what this looks like for you?

The audit is two weeks and ends in a costed roadmap you can execute with anyone — us included, or not.

Book a Audit