How we build recommendations

What data we collect

We collect structured responses from employees via a tap-based survey. Each survey covers one repetitive task and includes: the employee role and department, the task type, source and destination systems, the trigger event, frequency, time per occurrence, action type, and variability level.

There is an optional free-text notes field. No files, screens, emails, or system access are involved. Responses are linked to a company but not to a named individual in the report output.

Normalisation

Raw responses are first normalised — field values are mapped to a consistent internal vocabulary (e.g. "Gmail", "Outlook", and "email" are all mapped to the email system type). Weekly minutes are calculated from frequency and time-per-occurrence inputs. A variability score is assigned (1 = low, 2 = medium, 3 = high) which affects confidence scoring downstream.

Clustering

Normalised workflows are grouped into clusters using a composite key of task type, source system, and destination system. This means that if multiple employees report doing the same structural task — even if described differently — they are treated as a single cluster. ROI estimates are scaled by the number of people in each cluster.

Pattern matching

Each cluster is scored against a library of pre-vetted automation patterns. Scoring is field-level:

• 30 pts Source system match
• 25 pts Destination system match
• 25 pts Trigger type match
• 20 pts Action type match

A workflow must achieve a minimum match score of 40 to be considered for a recommendation. Partial credit is given for “other” field values to avoid false negatives.

Confidence scoring

The raw match score is adjusted by confidence rules:

• −30 points — highly variable task (hard to automate reliably)
• −15 points — somewhat variable task
• +10 points — high frequency task (5+ hours/week)
• +5 points — moderate frequency (2.5+ hours/week)
• −10 points — very low frequency (under 30 minutes/week)

Resulting scores are categorised as:

• High (75–100) — strong match, reliable starting point
• Medium (40–74) — solid lead, some investigation recommended
• Low (<40) — weak match, expert review recommended before acting

Recommendation classification

Each recommendation is classified into one of five types:

• AI Automation — tasks where AI can genuinely reduce manual effort (classification, summarisation, drafting)
• Workflow Automation — trigger-based automations using tools like Zapier, Make, or n8n
• Software Upgrade — replacing manual workarounds with purpose-built software
• Process Standardisation — the process needs cleaning up before automation is feasible
• Expert Review — confidence is too low to automate safely without specialist input

ROI estimation

ROI is estimated using a simple, transparent formula: weekly minutes saved × 52 weeks ÷ 60 × hourly rate. The hourly rate defaults to $50/hr but can be configured by the company admin to reflect actual employee cost including salary, benefits, and overhead. Time savings are capped at 85% of total task time to avoid overstating the benefit. Results are indicative estimates, not guarantees.

What we deliberately do not do

• ×Generate recommendations without a matching pattern in the vetted library
• ×Present low-confidence items as high-confidence
• ×Record screens, track keystrokes, or monitor activity
• ×Collect or process individual employee performance data
• ×Use LLMs to generate automation recommendations (all recommendations are pattern-matched)
• ×Guarantee implementation timelines or outcomes