Systems we have built.
Results we can defend.
These examples draw on hands-on operating experience designing, implementing and measuring AI systems across company knowledge, customer support, finance and compliance.
These case studies are anonymized composites of systems we built and operated. Names, products and identifying details are removed or generalized; the outcomes are real measured results, rounded. All interface visuals are original fictional mockups.
01 · Company knowledge
A company brain that answers with evidence
A distributed technology team was losing decisions to the places where work happens: chat threads, engineering tickets and meeting recordings. Finding out why something was decided meant asking whoever was there at the time. We built a governed knowledge system that collects those conversations automatically and turns them into answers people can check.
- Automated collectors for chat, ticketing and meeting transcripts - nightly archives plus a 15-minute signal scan, with no manual filing
- Governed knowledge sources with role-scoped access that fails closed, verified with a deliberately restricted account
- A weekly human review that decides what gets promoted from raw intake into curated company knowledge
- A query interface that answers in plain language with numbered citations back to the underlying records
- 3
- conversation systems ingested automatically
- 7
- knowledge sources with role-scoped access
- ~15 min
- tested restore time on fresh infrastructure
- 0
- manual collection steps left in the pipeline
Collection has run fully automated since mid-2026. The recovery figure comes from an actual restore drill on a fresh host, not an estimate.
People stay the editors: a weekly review curates what the company treats as true, and restricted sources are simply invisible to accounts without the right scope.
Chat, ticketing and meeting transcripts flow in on a schedule - nightly archives plus a 15-minute signal scan. Nobody files anything by hand.
02 · Customer support
Support that resolves, and knows when to stop
A consumer product with a high-volume support queue was spending skilled human hours on repetitive billing and access questions, while emotionally sensitive cases needed more attention than the queue allowed. We built an AI support agent that resolves routine work autonomously and routes everything sensitive to a person, with full context attached.
- Automatic classification of every incoming conversation by type, urgency and sensitivity
- Account and history lookup before any decision - subscription state, billing record, prior conversations
- Playbook-matched autonomous replies for routine cases, checked against policy before sending
- A hard routing rule: sensitive, emotional or high-urgency cases always go to a human
- ~40%
- of all conversations resolved autonomously
- ~70%
- autonomous resolution in the high-volume queue
- -31%
- median time to close
- -73%
- time to close at the 75th percentile
Measured over a seven-week pilot. Sensitive and high-urgency cases stayed human-routed by design throughout - the automation rate is a policy choice, not a ceiling the system hit.
The system knows its own boundary: anything emotional, complex or high-stakes is handed to a person with the account context and conversation history already assembled.
Incoming conversations are classified by type, urgency and sensitivity the moment they arrive.
03 · Finance operations
Revenue recognition with a control that checks the machine
A subscription business needed monthly revenue recognition across thousands of transactions, multiple currencies and annual cohorts - work that was slow to prepare and hard to verify by hand. We encoded the accounting policy into a workflow, then built a second, independent calculation whose only job is to catch the first one being wrong.
- A policy-controlled recognition workflow: cohort treatment, currency conversion and exclusions applied the same way every month
- An independent recalculation built on exact daily data as a standing reconciliation control
- Exception thresholds that flag any variance beyond an agreed tolerance for investigation
- A human approval gate - the system prepares the schedule and draft entries, a person reviews and posts
- <2%
- variance between primary run and independent control
- 4
- consecutive months tested and reconciled
- 100%
- of postings pass through human approval
Reconciliation held within the 2% tolerance in every tested month; the final tested month closed at roughly 1.5%, explained by timing and currency differences.
The agent never posts to the ledger autonomously. It prepares, reconciles and explains - a finance owner approves. Variances above threshold stop the workflow.
Subscription transactions stream in from the billing system - thousands per month, across currencies and plan types.
04 · Compliance operations
A compliance evidence backlog cleared in one working day
Ahead of external certification audits, a technology business faced a backlog of open evidence requests across two security frameworks - each needing a document that reflected how the company actually operates. We built an evidence pipeline that triages the backlog, drafts documents from real company records, and submits them through the compliance platform with review built in.
- Backlog triage that maps every evidence request to its owner, control and risk
- Document generation grounded in real company records - policies, registers, test logs and file notes
- An upload-and-submit loop through the compliance platform's API, with rate-limited, verifiable status transitions
- A risk register linked to specific controls, so evidence, risk and control stay one connected picture
- 66
- evidence requests triaged in one pass
- 51
- documents produced and submitted in one working day
- 22 → 67
- framework A documents passing, of 81
- 34 → 38
- framework B documents passing, of 42
Counts are from the platform's own status dashboard on the day. Seven mapped risk scenarios and a later clean external audit came from the broader compliance program this system supported - not from automation alone.
Every document is reviewed before submission and the compliance platform remains the system of record. Auditors only ever see human-approved evidence.
Sixty-six open evidence requests are mapped to owners, controls and risks in a single pass.
Your workflow next
Every one of these started as one frustrating process.
Bring the workflow that wastes the most of your team's time. In a first working session we will map what happens today and whether a small AI system is worth testing.
hello@denkwerk.aiOr try the live Company Brain example