Internal strategy review
May 2026 · v1
The next three to five years
The future of app engineering.
Three lenses on a practice being restructured by AI: the engineer, the practice, the firm.
Workforce strategy · Confidential
01 / 35
Thesis
02 / 35
In one page

AI does not slowly change app engineering. It restructures it.

Unit of work
From developer-hour to verified outcome.
The atomic unit a firm sells, staffs, and measures shifts from a billable hour of human authorship to a verified deliverable produced by a human-plus-agent pod.
The gap
Almost no large practice is built for that unit.
Operating model, pricing, comp, and tooling are still optimized for billable hours. Realized productivity is a small fraction of what is technically available.
The prize
Winners rebuild the system before pricing collapses.
The firms that migrate delivery, commercials, and talent first capture a multi-year window before T&M margins compress structurally.
Workforce strategy
02 / 35
How to read this deck
03 / 35
Three lenses, one story

We tell the story at three altitudes.

Act I — Lens 01
The engineer.
The person at the keyboard. Where the daily work changes shape first.
~1 engineer · slides 04 – 14
Act II — Lens 02
The practice.
Ten thousand engineers, hundreds of accounts. Where individual changes compound — or cancel.
~10,000 engineers · slides 15 – 24
Act III — Lens 03
The organization.
The services firm. Where commercial model, capital, and competitive position get rewritten.
The firm · slides 25 – 34
Each act follows the same arc: future state → today's gap → frictions → missed moves → competitive view → our strategic answer.
Workforce strategy
03 / 35
Act I
04 / 35
Act I · The engineer
I.
The engineer.
What the person at the keyboard does in 2029 — and why nothing else makes sense without starting there.
05 — 07
Future state & role shift
08 — 11
The gap & the frictions
12
Competitive view
13 — 14
Our strategic answer
Act I · Future state
05 / 35
The engineer in 2029

Five qualities define the engineer who is in demand.

01 — Specification
Writes intent precisely.
Specifies units of work in a form an agent can execute well, with explicit success criteria, constraints, and acceptance tests.
02 — Decomposition
Cuts problems for agents.
Knows the granularity at which agents succeed, and where a human seam between agent steps is required.
03 — Verification
Reviews machine output fast.
Runs a deliberate verification workflow — automated harnesses plus targeted human review — at the throughput agents demand.
04 — Architecture
Owns the seams agents cannot.
Integration boundaries, data semantics, performance edges, security model, and the question of what to build at all.
05 — Trust calibration
Knows when to look closely.
Has a calibrated sense for where agent output is reliable and where it is not, and budgets verification time accordingly.
Act I — The engineer
05 / 35
Act I · Future state
06 / 35
A representative day · 2029

Hours shift from authoring to specifying and verifying.

07 : 00
Reviews overnight agent work. Three branches produced unattended; two pass the harness, one is rejected with a note.
09 : 00
Specification session. Writes acceptance tests and constraints for the next two units of work before any code is touched.
11 : 00
Hands-on integration. Wires a payments boundary by hand — ambiguity is highest, agent error most expensive.
14 : 00
Verification sprint. Drives the eval harness against the day's agent output; triages, writes new tests, lands branches.
16 : 00
Pod review. Walks through architecture decisions with three peers — what to build, not how to type it.
Time allocation · illustrative
Authoring code
~15%
Specifying work
~25%
Verifying output
~35%
Architecture & review
~25%
Directional. Distribution varies by domain; ranges, not points.
Act I — The engineer
06 / 35
Act I · Role shift
07 / 35
The role shift

From author to orchestrator.

2024 · Author
The engineer produces code. The IDE assists.
Primary output
Lines of code, by hand.
Leverage source
Personal speed, library knowledge.
Time on review
Minutes per day on peer PRs.
Identity
"I write software."
Measured by
Tickets closed, PRs merged.
→
2029 · Orchestrator
The engineer produces intent & judgement. Agents produce the code.
Primary output
Specifications, verifications, architectural decisions.
Leverage source
Fleet of agents, evaluation harness.
Time on review
Hours per day on machine output.
Identity
"I design systems that ship themselves."
Measured by
Verified outcomes shipped.
Act I — The engineer
07 / 35
Act I · The gap
08 / 35
Today's gap · 2026

Most engineers use AI in a fundamentally 2022 mode.

~85%
use AI primarily as smarter autocomplete inside their editor.
<15%
run a genuine agentic workflow — specify, dispatch, verify — as their default mode.
~1 in 10
can articulate a deliberate verification workflow for agent output.
0
hours of formal training, today, on writing specifications agents can execute well.
Internal pulse, Q1 2026, n ≈ 2,100. Directional.
Act I — The engineer
08 / 35
Act I · The gap
09 / 35
The effectiveness cliff

Output per engineer is no longer normally distributed.

Verified output per engineer · indexed to 2023 = 100 · illustrative
BOTTOM DECILE MEDIAN TOP DECILE OUTPUT INDEX Slight decline — fighting tools 3 – 5× baseline output 2023 baseline flat 2026 2023
Illustrative shape derived from internal pulse data and external research. Y-axis is indexed; absolute values intentionally suppressed.
The shape
Bimodal and pulling apart every quarter.
The implication
The top decile is several times more productive than the median. Best practices are not yet portable. This is a people-management problem first, a tooling problem second.
Act I — The engineer
09 / 35
Act I · Frictions
10 / 35
Five frictions · engineer level

Where individual engineers get stuck.

01
Tool sprawl.
Four to six AI surfaces per engineer, none consolidated. Switching tax compounds across the day.
02
The trust problem.
Engineers either over-trust agent output and ship defects, or re-verify by hand and erase the speed gain.
03
Evaluation.
No credible signal of whether agent code is good. Reviewers fall back to surface heuristics.
04
Identity.
Many engineers signed up to write code, not supervise machines. Morale and retention take the hit quietly.
05
Learning curve.
The best workflows need deliberate practice. Without ring-fenced time, the cliff steepens.
Act I — The engineer
10 / 35
Act I · Missed moves
11 / 35
Missed moves · how the industry is responding

Three patterns of suboptimal response we see today.

Pattern A
Procurement over workflow.
Treating AI tools as a license decision. Buy seats, declare adoption, move on. Engineers receive credentials, not a redesigned way of working.
→ License utilization rises. Output does not.
Pattern B
Wrong metric.
Measuring adoption by seat count, prompt count, or completion-acceptance rate, rather than verified output shipped. Produces the wrong incentives across the org.
→ Theater of adoption. No correlated outcome.
Pattern C
Workflow as hobby.
Leaving the design of agent workflows to individual engineers in their spare time, instead of as a first-class engineering investment with named owners and measurable goals.
→ Best practices stay siloed in the top decile.
Act I — The engineer
11 / 35
Act I · Competitive view
12 / 35
Competitive view · at the engineer level

Three competitors are pulling at our engineers.

Competitor What they do for engineers Where it threatens us
Large peer firms Match tooling spend; broad training programs; centralized copilots. Comparable comp bands to ours. Same engineers we want; same playbook. Whoever moves first on workflow and comp wins.
AI-native upstarts Ten-engineer teams bidding work that used to require fifty. Pay top-decile engineers 2 – 3× market. No legacy delivery org. Bleed our top decile. Set the buyer's mental anchor for "modern" delivery in their favor.
Client in-house teams Pulling our best engineers into product roles. Own the codebase, so agent leverage compounds inside their walls. Long-term: clients need us less if they internalize agent operations on their own platforms.
Us · status quo Broad tooling rollout. Adoption metrics. Training as e-learning. Comp unchanged from 2023 bands. Caught in the middle. Top decile leaves to upstarts; median commoditizes against peers.
Act I — The engineer
12 / 35
Act I · Our strategic answer
13 / 35
Our strategic answer · engineer level

Build & pay for the T-shaped agent operator.

HORIZONTAL BAR Multi-agent orchestration · specification · verification VERTICAL BAR Deep specialty — Distributed systems — Security architecture — Vertical data model — Regulated domains Judgement that does not commoditize.
Why this shape
The full-stack generalist we hired against for a decade is the wrong target. Agents flatten generalism; depth plus orchestration compounds.
Grade both bars
Pay for vertical depth and horizontal fluency, separately. Top comp band reserved for engineers who carry both.
Training implication
Two distinct curricula, two distinct certifications, two distinct career ladders that intersect.
Act I — The engineer
13 / 35
Act I · Moves
14 / 35
Four concrete moves · next 12 months

What we change for engineers.

01
Agent-operator certification, three levels.
Gated by verified output on real work, not by attendance. Required for promotion past senior engineer. Re-certified annually.
02
200 hours of paid learning time per engineer per year.
Ring-fenced, scheduled, not at the manager's discretion. Logged against the certification curriculum. Funded centrally.
03
Comp bands rebuilt around verified output.
Top band reserved for top-decile agent operators who also carry a vertical bar. Multi-year transition, communicated openly.
04
A central golden-path tools team.
Small, senior, well-funded. Owns the standard agent workflow so individual engineers do not each invent their own. Treats internal engineers as customers.
Act I — The engineer
14 / 35
Act II
15 / 35
Act II · The practice
II.
The practice.
Ten thousand engineers. Hundreds of accounts. Where individual changes either compound — or cancel out.
16 — 17
Future state & model shift
18 — 21
The gap & the frictions
22
Competitive view
23 — 24
Our strategic answer
Act II · Future state
16 / 35
The practice in 2029

Pods of humans and agents. A verification layer at the heart.

Unit of staffing
The pod, not the person.
A pod is 4 – 7 humans plus a persistent agent fleet. Carries the scope a 12 – 20 person team carried in 2024. Scheduled against outcomes, not seats.
Throughput governor
The verification layer.
A dedicated capability — senior engineers plus eval infrastructure — whose throughput sets the pod's ceiling. Not optional, not a side activity.
Capability centers
Specialized, not generic.
No longer staff-augmentation pools. Each center owns a vertical, an evaluation discipline, or a platform layer. Headcount down; specialization up.
Delivery management
Pod manager replaces resource manager.
Manages throughput and quality of a small number of pods, not a roster of 30 named individuals. New training, new tooling, new comp.
Primary metric
Pod throughput, not utilization.
Utilization is no longer the lever. Verified outcomes shipped per pod per quarter is the lever. Every system above this layer follows.
Act II — The practice
16 / 35
Act II · Model shift
17 / 35
The operating model shift

From resource-loaded plan to throughput-managed pod.

Today · Resource-loaded plan
Staff N engineers. Bill against hours. Measure utilization.
Unit of staffing
A named person, on a rate card.
Primary lever
Utilization rate.
Sales motion
Sell hours; rate-card discussion.
Staffing tool
Roster, skills, availability.
QA model
Sample-based human review.
SOW shape
Engineer-month, T&M.
→
2029 · Throughput-managed pod
Deploy a pod against an outcome. Bill against delivery. Measure throughput & quality.
Unit of staffing
A pod with declared capacity & specialty.
Primary lever
Pod throughput & verification capacity.
Sales motion
Sell outcomes; pricing on value.
Staffing tool
Pod catalogue, eval-passing capacity.
QA model
Automated eval harness + targeted human review.
SOW shape
Outcome SOW, gainshare, milestoned.
Act II — The practice
17 / 35
Act II · The gap
18 / 35
Today's gap · the practice

Every system in the practice is designed for an obsolescing unit of work.

A
Utilization drives staffing.
A pod that ships in half the time looks like under-utilization in our reporting.
B
The staffing tool cannot represent a pod.
It models people. Mixed human-agent capacity has no native data model.
C
PMs run resource-loaded plans.
Trained on Gantt charts, not throughput. New skill set; old job description.
D
SOWs price an engineer-month.
The commercial unit is the bottleneck. Outcome SOWs require a different finance posture.
E
QA is sample-based human review.
It does not scale to agent throughput. The eval harness is the new QA function.
F
Career ladders reward seniority of person, not pod outcome.
High-throughput pod operators have no obvious next rung.
Act II — The practice
18 / 35
Act II · The gap
19 / 35
The productivity paradox

Capability is growing exponentially. Realized productivity is growing linearly.

Agent capability vs realized practice productivity · indexed · 2023 = 100 · illustrative
THE GAP = THE PRIZE 2023 2024 2025 2026 INDEX Agent capability Realized practice productivity
Shape from external research and internal pulse. Axis indexed; absolute values suppressed by design.
The gap
~20% of theoretical productivity gain has been captured by the median practice today.
The reason
Operating-model problem, not a tooling problem. Closing it is structural, not procurement.
The prize
The shaded area is the size of the prize for the first firm at scale to close it.
Act II — The practice
19 / 35
Act II · Frictions
20 / 35
Five frictions · practice level

Where the practice stalls.

01
Evaluation infrastructure is missing.
We cannot tell whether pod A is twice as productive as pod B because we have no uniform measure of pod output. Every comparison is anecdote.
02
Client codebases are not agent-ready.
Most engagements inherit codebases that cap agent throughput. The readiness work is real, but invisible in current SOWs.
03
IP & data boundaries are immature.
Agents need context; clients are sensitive. Our governance is improvised. This is becoming a deal-killer in regulated verticals.
04
Pricing inversion.
When a pod ships in half the time on a T&M contract, our revenue halves. Our SOW shape caps our upside on the work we do best.
05
Manager capability.
Most delivery managers have never run a pod. They do not yet know what good looks like, which means they cannot coach or staff for it.
Act II — The practice
20 / 35
Act II · Missed moves
21 / 35
Missed moves at scale

Three patterns of suboptimal response at the practice level.

Pattern A
Uniform rollout.
Deploying AI tooling identically across ten thousand engineers instead of concentrating investment in pods that can demonstrably absorb it. The cliff problem, repeated at scale.
→ Even thinness everywhere. Depth nowhere.
Pattern B
Adoption-as-output.
Building an evaluation function that grades AI adoption metrics rather than verified delivery. Multiplies the wrong incentive across the entire practice.
→ Adoption dashboards green. Delivery flat.
Pattern C
Codebase-readiness as client problem.
Treating the work of making a codebase agent-ready as the client's responsibility, rather than as a paid service line we deliver before — and during — the build.
→ Revenue left on the table. Throughput stays capped.
Act II — The practice
21 / 35
Act II · Competitive view
22 / 35
Competitive view · at the practice level

Three categories of competitor, three different threats.

Competitor What they do at the practice level Where it threatens us
Large peer firms Roughly our playbook: training, partnerships, internal copilots. Same legacy operating model. Same SOW shapes. The race is who rebuilds the operating model first. First mover wins a multi-year window.
AI-native delivery shops Pod-based commercials on small to mid scope. Demonstrably faster on the work they take. Win on brand and speed. Anchor the buyer's mental model of "modern" delivery. Take the entry deals into our accounts.
Client in-house teams Insource the work where agent leverage is highest and they own the codebase. Outsource the rest. Slowly hollow our scope inside major accounts. The most strategic threat over five years.
Us · status quo Tooling deployed broadly. Utilization still the metric. SOWs unchanged. AI is a delivery enhancement story. Lose to upstarts on speed; lose to peers on cost; lose to in-house on scope.
Act II — The practice
22 / 35
Act II · Our strategic answer
23 / 35
Our strategic answer · practice level

The agentic delivery factory, in three layers.

Layer 03
Platform layer
Owns the golden-path agent workflow, the evaluation harness, and the safety & compliance perimeter. Built once, used by every pod.
"Built once. Used by every pod."
Layer 02 · The heart
Verification layer
Senior engineers plus eval infrastructure. Throughput here governs the whole system. Run as a profit center, with its own metrics & investment plan.
"Throughput here = throughput everywhere."
Layer 01
Pod layer
4 – 7 humans plus persistent agent fleet. Organized by capability, not by account. Scheduled against outcomes. Owns delivery and quality.
"The pod is the unit. Not the person."
The factory is what we sell. The factory is what we improve. Every other practice investment serves it.
Act II — The practice
23 / 35
Act II · Moves
24 / 35
Four concrete moves · next 18 months

What we change at the practice.

01
Replace utilization with pod throughput as the primary delivery metric.
In the staffing tool, the financial system, the delivery review, the QBR. The change is in the systems, not just the slides.
02
Stand up a central evaluation team and platform.
Without it, none of the other moves are measurable. Funded centrally; staffed with our most senior engineers.
03
Build & require a paid codebase-readiness service line.
Required on outcome-priced engagements. Captures revenue today; uncaps throughput later.
04
Retrain the top 500 delivery managers on pod management.
Re-assign or exit the ones who cannot run pods within a year. The kindest version of an unkind constraint.
Act II — The practice
24 / 35
Act III
25 / 35
Act III · The organization
III.
The organization.
The services firm. Where commercial model, capital allocation, and competitive position get rewritten.
26 — 27
Future state & commercial shift
28 — 31
The gap & the frictions
32
Competitive view
33 — 34
Our strategic answer
Act III · Future state
26 / 35
The services firm in 2029

A platform business with a delivery arm — not a delivery business with a platform attached.

Revenue mix · 2024 vs 2029 · illustrative
T&M · 2024
~92%
T&M · 2029
~25%
Outcome / gainshare · 2024
~6%
Outcome / gainshare · 2029
~50%
Platform / licensing · 2024
~2%
Platform / licensing · 2029
~25%
Aspirational mix at the firm level. Illustrative, not a forecast.
Margin profile
Blended margin higher than today — but only in platform and outcome segments. T&M margin structurally lower.
Center of gravity
Platform & evaluation are the firm's strategic core. Delivery is the customer.
Customer mix
Platform sold to clients and to delivery partners — including former competitors.
Act III — The organization
26 / 35
Act III · Commercial shift
27 / 35
The commercial model shift

From bodies to outcomes.

Today · Bodies
90%+ of revenue sold as headcount on a rate card.
Unit sold
Engineer-month, by skill & geography.
Primary lever
Utilization.
Risk posture
Time-and-materials. Client bears delivery risk.
Underwriting
None required.
Moat
Scale of supply & geography.
→
2029 · Outcomes
Dominant unit is a verified deliverable, priced against value.
Unit sold
A verified outcome, milestoned.
Primary lever
Pod throughput & platform leverage.
Risk posture
Outcome & gainshare. We bear delivery risk.
Underwriting
Required. Strategic moat in itself.
Moat
Platform, evals, regulated-vertical depth.
Act III — The organization
27 / 35
Act III · The gap
28 / 35
Today's gap · the firm

Every lever above delivery is still set for the old unit.

A
Revenue is heavily T&M.
Concentrated on the unit most exposed to AI-driven cycle-time compression.
B
Sales comp is booked revenue, not gross margin.
Sellers are paid to fill seats, not to land outcome SOWs at premium margin.
C
Partnerships are tooling resale.
We resell other people's models. We do not co-design the offering.
D
M&A history is geographic.
Labor-arbitrage targets, not capability or platform targets. Path-dependent.
E
Brand is "reliable scale."
Not "technical edge." With buyers who pay premiums for edge, we are not on the shortlist.
F
Capital allocation favors tooling licenses.
Most internal R&D spend rents capability. Almost none builds proprietary platform.
Each one is fixable. None fixes itself.
Act III — The organization
28 / 35
Act III · The gap
29 / 35
The margin scissor

Two curves close on T&M margin in the next 24 – 36 months.

T&M rate vs platform investment cost · indexed · illustrative
SCISSOR CLOSES NOW +12 MO +24 MO +36 MO T&M effective rate (cost-recovery) Platform & evaluation investment cost
Indexed. The exact crossover varies by segment; the shape does not.
Top blade
T&M rates flat to declining as the engineer-hour produces less of the customer's value.
Bottom blade
Platform & evaluation investment rising to fund the factory and the moat.
Implication
Firms without meaningful revenue out of T&M by the crossover face a step-change in margin compression.
Act III — The organization
29 / 35
Act III · Frictions
30 / 35
Five frictions · firm level

Where the firm gets stuck.

01
Outcome underwriting.
Finance does not yet know how to price and reserve against outcome SOWs. Without this capability, sales cannot sign them.
02
Capital allocation.
Most internal R&D rents capability from model providers rather than building proprietary platform. Strategically thin.
03
Partnership posture.
We are a reseller of models, not a co-designer of an offering. The economics of reselling are not the economics of a moat.
04
M&A logic.
Our pipeline is still labor-cost-arbitrage targets when it should be capability and platform targets in two or three specific segments.
05
Brand.
We are not known for technical edge with the buyers who pay premiums for it. The brand asset is reliable scale — the wrong moat for this decade.
Act III — The organization
30 / 35
Act III · Missed moves
31 / 35
Missed moves at the firm

Three patterns of suboptimal response across the industry.

Pattern A
AI as enhancement, not offering.
Treating AI as a delivery enhancement story for existing accounts. The bigger prize is rewriting the offering portfolio, not garnishing it.
→ Defending the old book; never landing the new one.
Pattern B
Ceding brand to upstarts.
Letting AI-native upstarts define what "fast" and "modern" look like in the buyer's mind without contesting it. By the time we do, the anchor is set.
→ We are the safe choice. Not the right one.
Pattern C
Platform as cost center.
Keeping platform & evaluation reporting into delivery as overhead, rather than as profit centers with their own P&L and outside-customer obligations.
→ Underinvested. Slow. Captive to delivery's needs.
Act III — The organization
31 / 35
Act III · Competitive view
32 / 35
Competitive view · at the firm level

Three categories. Each strong somewhere. None doing all of it.

Competitor Strength Weakness Where we beat them
Large peer firms Scale, account access, regulated footprint. Same legacy operating model and pricing as us. Move first on commercial model. Reset the anchor for what large-firm modern looks like.
AI-native upstarts Pod-native ways of working. Speed brand. No scale, no regulated-vertical depth, no balance sheet to underwrite outcomes. Outcome SOWs at scale, regulated verticals, multi-year transformation deals they cannot bid.
Client in-house teams Own the codebase. Compounding agent leverage. Cannot platform across companies; cannot hire all the depth they need. Sell platform & evaluation back to them. Be inside their workflow, not outside it.
Our position Scale + regulated depth + balance sheet to underwrite. Operating model and brand still set for the prior unit. Be the only firm doing all three: factory, outcome model, platform.
Act III — The organization
32 / 35
Act III · Our strategic answer
33 / 35
Our strategic answer · firm level

Become an outcome platform.

Layer 03 · The moat
The platform
Evaluation, golden-path agent workflows, safety & compliance perimeter. Operated by us, sold by us — to clients and to delivery partners.
"Sells to the people who used to compete with us."
Layer 02 · The revenue
Outcome-priced delivery
Pods on top of the platform. Outcome & gainshare SOWs underwritten by the firm. The premium-margin core of the business.
"Underwriting is a moat AI-natives cannot match."
Layer 01 · The transition
T&M tail
The classic book. Funds the migration. Converts to outcome as accounts mature. Managed for cash, not growth.
"The bridge — not the destination."
Each layer funds the layer above it. The platform is the moat. The outcome offering is the revenue. T&M is the bridge.
Act III — The organization
33 / 35
Act III · Moves
34 / 35
Five firm-level bets · next 24 months

What we change at the firm.

01
Carve out platform & evaluation as a standalone business unit.
Own P&L. Mandate to sell externally — to clients and to other delivery partners. Reports into the CEO, not into delivery.
02
Restructure sales comp to gross margin and outcome attainment.
Bookings remain a guardrail. Margin and outcome retention become the levers. Communicated 12 months ahead of the change.
03
Redirect M&A toward 2 – 3 platform and evaluation acquisitions.
Inside 18 months. Capability and platform targets, not labor-arbitrage geographies. Board mandate.
04
Sign one anchor outcome-priced deal per major vertical.
Establishes reference pricing, proves the underwriting capability, creates the case study the sales force will need.
05
Fund a brand campaign on outcome partnership in regulated verticals.
Reposition from "reliable scale" to "the outcome partner for mission-critical app engineering." Where our footprint compounds.
Act III — The organization
34 / 35
Synthesis
35 / 35
The three answers, stacked

The strategy is one thing, told three times.

Lens
Engineer · Act I
Practice · Act II
Organization · Act III
Future state
T-shaped agent operator. Depth + orchestration.
Agentic delivery factory. Pods, platform, verification.
Outcome platform. Platform · outcomes · T&M tail.
Primary unit
Verified output by a person + agents.
Pod throughput.
Outcome & platform revenue.
Key move
Certify & pay both bars of the T.
Replace utilization with throughput as the metric.
Carve out platform as standalone BU with external mandate.
If we don't
Top decile leaves to upstarts; median commoditizes.
Lose to upstarts on speed, to peers on cost, to in-house on scope.
Margin scissor closes underneath us.
The answers are deliberately stacked: the engineer answer enables the practice answer, and the practice answer enables the firm answer. Doing one without the others is wasted motion. Doing all three, ahead of peers, is the strategy.
Workforce strategy · End
35 / 35