Skip to content

Ottawa · AI adoption

Your team could be getting a quarter of its week back.

That is what controlled trials keep finding when people are given the right AI tools for the right tasks. The hard part is knowing which tasks, in your organisation, with your data. That is the part we do.

Measured in controlled trials
More work completed, 4,867 developers [1]
+26%
Faster delivery, 758 consultants [2]
+25%
Quality rating of the work produced [2]
+40%
Cases resolved per hour, 5,179 agents [3]
+14%

Randomized trials, not vendor estimates. Papers listed at the foot of this page.

What the research actually shows

The gains are real, and they are largest for your newest staff.

Both of the big workplace experiments found the same thing: the people who improved most were the least experienced. AI narrows the gap between your strongest performer and everybody else, which matters more in a team of nine than in a team of nine hundred.

More work completed4,867 developers · 3 trials
+26%
Faster task completion758 consultants
+25%
More cases resolved per hour5,179 support agents
+14%

Effects from randomized controlled trials. Bars scaled to 40%. Full citations below.

Who the gains go to

Novice and lower-skilled staffcustomer support · 5,179 agents
+34%
Already-experienced staffsame study, same period
~0%

The 14% average hides this split [3]. In a team of nine, the 34% is the number that changes your week.

Across three randomized trials at Microsoft, Accenture and a Fortune 100 firm, 4,867 developers given an AI assistant completed 26% more tasks. [1]

758 consultants in a controlled experiment finished 25% faster and produced work rated over 40% better than the group working without AI. [2]

5,179 customer support agents resolved 14% more issues per hour with an AI assistant, measured against their own prior performance. [3]

Where this lands in your week

Six places it pays, in most organisations.

The returns show up in repetitive, verifiable work, the things already costing your team hours and nobody’s favourite part of the job.

Writing that follows a pattern

Grant applications, funder reports, proposals, quotes, listings. The research and the judgement stay yours. The draft stops taking an evening.

Fastest to prove

Paper turned into data

Invoices, receipts, intake forms, referrals. Read once, filed correctly, with anything uncertain put in front of a person instead of guessed at.

Highest hours saved

The first reply

Enquiries triaged and answered as drafts your team approves. The after-hours message that used to wait until Monday gets handled.

Most visible to clients

Knowing what you already know

Policies, past files, meeting notes and procedures made answerable. New staff stop interrupting senior people to find things.

Best for small teams

Reporting that writes itself

The monthly numbers assembled and summarised from your own systems, with the figures pulled from your data rather than produced by a model.

Recurring time back

Meetings that produce actions

Decisions, owners and dates extracted and pushed into whatever you already use, so the follow-through stops depending on who took notes.

Cheapest to start

Worked examples

What this looks like as arithmetic.

Three illustrations, with the sums shown so you can argue with them. These are hypothetical organisations, not clients of ours. Each links to a full write-up. The per-task savings are set well below the trial effects above. A controlled experiment is evidence that something works, rather than a promise about your week.

Funder reporting

A nine-person charity

Today

Reports a year6
Director-days each, assembling and writing2
Days a year12

With the change

Numbers pulled from the finance systemautomatic
Director-days each, editing a draft0.5
Days a year3

≈ 9 director-days returned

Stays with a person: Every figure comes from their own finance system. The model never produces a number, and the director signs each report before it leaves.

Read the full example, how it is built, and what we would not automate →

Quoting

A four-person trades business

Today

Quotes a month40
Minutes each, written in the evening25
Hours a month17

With the change

Draft built from site photos and notesautomatic
Minutes each, checking and sending8
Hours a month5

≈ 12 hours a month, and quotes go same-day

Stays with a person: Prices come from their own rate card, never from the model. The owner approves every figure before it reaches a customer.

Read the full example, how it is built, and what we would not automate →

Intake and referrals

A twelve-person clinic

Today

Documents a week30
Minutes each, keyed by hand10
Hours a week5

With the change

Read and filed, uncertain ones queuedautomatic
Minutes each, reviewing the queue3
Hours a week1.5

≈ 3.5 hours a week back to clinical staff

Stays with a person: Anything the extractor is unsure of goes to a person rather than being guessed. What may be sent to a model is written down before a single document is processed.

Read the full example, how it is built, and what we would not automate →

Where it does not pay

The same research shows where this goes wrong.

Telling you what to leave alone is part of the job, and it is the part that keeps you out of trouble.

WhereWhat the evidence says
Experienced people on familiar workOne trial of seasoned open-source developers found AI made them 19% slower on real tasks. The gains concentrate where the work is new or unfamiliar, which is why we measure before and after rather than assume. [4]
Anything with a number that mattersA model asked for a figure will produce one. In a funder report or a board pack the numbers come from your systems, never from a model.
Pilots with no ownerMIT found 95% of enterprise AI pilots produced no measurable profit-and-loss impact. The failures were strategy, data readiness and ownership, not the technology. [5]

How we work with you

Find the hours, build the thing, keep it running.

1 · Readiness review

We map your week, find where these numbers apply to your actual work, and hand you a written plan with what each change is worth. You keep it either way.

Grant-eligible in Ontario [7]

2 · Build

We take the two or three with the clearest return and build them into your existing tools, measured against what the work cost before.

Measured before and after

3 · Run and train

We keep it working, write the rules for what staff may put into these tools, and train people on the parts that go wrong.

Where 95% of pilots fail [5]

Why us

We run this ourselves, daily.

An Ottawa news site that publishes every day through a pipeline we built, with a corrections queue a person signs off. A payments platform with 46 businesses on it. We write the code, take the support calls, and fix it when it gets something wrong, so the advice comes from operating these systems rather than from reading about them.

See the work →

Also built here

Web applications · mobile apps · payment integrations · a bilingual, accessibility-first site for a healthcare institution · technical SEO

Start with the hours you would get back.

One call. Tell us what your team spends its week on and we will tell you which parts of it these numbers apply to, and whether there is anything here worth paying for yet.

Every figure behind the argument

Sources.

  1. Cui, Demirer, Jaffe, Musolff, Peng & Salz, “The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers” (February 2025). 26.08% increase in completed tasks (SE 10.3%) across 4,867 developers at Microsoft, Accenture and a Fortune 100 company. Less experienced developers showed greater gains. link ↗
  2. Dell’Acqua et al., “Navigating the Jagged Technological Frontier”, Harvard Business School / BCG. 758 consultants; 12.2% more tasks, 25.1% faster, outputs rated over 40% higher in quality. The same study found AI users more likely to be wrong on tasks beyond the model’s capability.
  3. Brynjolfsson, Li & Raymond, “Generative AI at Work”, NBER Working Paper 31161. Issues resolved per hour rose 14% on average across 5,179 customer support agents, a 34% improvement for novice and lower-skilled workers, with minimal impact on experienced and highly skilled ones. A later draft reports 15% average on 5,172 agents. Working-paper versions; not peer reviewed. link ↗
  4. A randomized trial of experienced open-source developers found AI tools increased completion time by 19% across 246 real tasks. Reported via secondary summary; we cite it because the counter-evidence matters as much as the gains.
  5. MIT NANDA, “The GenAI Divide: State of AI in Business” (2025). 95% of enterprise GenAI pilots produced no measurable profit-and-loss impact. link ↗
  6. Thryv small-business survey, April 2026, n=561. 66% of small and mid-sized businesses use AI; 70% say they need more training to use it effectively. link ↗
  7. Ontario Digital Competence Centre, delivered by the Ontario Centre of Innovation. Matched planning grant to $15,000; implementation grant to $50,000. Confirm current terms and eligibility with OCI. link ↗