Bottom-Up vs Top-Down AI Adoption: Enterprise Guide 2026

The 50-People Problem with Top-Down AI Assessments

The standard Tier-1 strategy playbook for an enterprise AI assessment looks the same whether the deck has McKinsey, BCG, or Bain on the cover. Interview 30–50 executives. Triangulate against an industry benchmark. Produce 20–30 prioritised use cases. Bind it into a 200-page deck. Bill seven figures. Move on.

That process has a structural problem the slideware never mentions: the other 850 employees do the actual workflows. The real AI ROI lives in the customer care queue at 11am on a Tuesday, in the agent onboarding flow that takes six manual handoffs, in the dealer dashboard that nobody updates because the export is broken. None of this is in the CFO's head. None of it is in the COO's head. It is in the heads of the 850 people the assessment never spoke to.

This is the argument for bottom-up AI adoption: the structural case for changing where AI strategy work actually happens inside an enterprise. If you are reading this, you have probably scoped a top-down assessment, received the proposal, and started wondering whether 50 executive interviews are really going to surface the inefficiencies your front-line teams quietly route around every day. The honest answer is no — and this post explains what to do instead, when bottom-up wins, when top-down still wins, and what a credible alternative engagement actually looks like.

What "Bottom-Up AI Adoption" Actually Means

The term bottom-up AI adoption gets used loosely. Sometimes people mean "grassroots AI" — individual employees using ChatGPT in the browser without IT approval. Sometimes they mean "shadow IT AI" — an analyst quietly building a Claude pipeline on a corporate laptop. Sometimes they mean "departmental AI" — one team buys Copilot, another buys Glean. Those are all real phenomena, and most large enterprises now have all three happening at the same time.

That is not what we mean here. In the context of an enterprise AI program, bottom-up AI adoption is a structured discovery method that surfaces opportunities and constraints from every layer of the organisation — not just the executive cohort. It is the opposite of "interview the C-suite and extrapolate". It is closer to "interview enough of the workforce, in enough channels, that the deck writes itself from real evidence."

Three distinctions matter:

• Structured, not grassroots. Bottom-up here is a deliberately designed discovery process, run by an external team with a methodology — not an emergent pattern of individual tool use. It produces a deliverable, KPIs, and a roadmap.
• Cross-layer, not just front-line. Bottom-up is not "only talk to the analysts and ignore the CFO." It reaches every layer. The CFO still gets interviewed. So do their direct reports. So do their direct reports. The point is reach, not inversion.
• Operational, not aspirational. The output is workflows, not themes. "AI in finance" is a theme. "Accounts payable invoice triage takes 14 minutes per invoice in two ERPs that don't share a vendor master" is a workflow. Bottom-up surfaces the second kind.

Bottom-up AI adoption is not a religion. It is a methodology choice that fits a specific problem shape: where the inefficiencies live in operational detail that doesn't roll up to a board pack. Most AI efficiency programs fit that shape. Most strategic-portfolio questions don't.

The Three-Channel Discovery Method

A credible bottom-up AI adoption engagement runs on three discovery channels in parallel. Each surfaces a different kind of signal. The synthesis happens when you triangulate across all three.

Channel 1: Deep Interviews (30–50 key roles, 60–90 min)

Not a survey. Real, hour-long conversations with people who own the actual workflows — not just the executives who report on them. The target list is built by mapping the org chart against revenue, headcount, and cost concentration, then picking the 30–50 roles where a workflow improvement would move a real number. Some are SVPs. Some are individual contributors. The deciding question is not seniority; it is "does this person know how the work actually flows?"

The output is a structured set of opportunity hypotheses, each grounded in a specific workflow narrated by the person who runs it. The interviewer is looking for friction signals: handoffs, swivel-chair work, manual reconciliation, exceptions that route to a senior person, ticket queues that quietly grow. Each one is a candidate for an AI efficiency opportunity.

Channel 2: AI-Driven Async Surveys (5–7 min, conversational, multilingual)

Where interviews give depth on 30–50 roles, async surveys give breadth across hundreds or thousands. The "AI-driven" part is what makes them work: a static 40-question form gets 8% response rates and zero useful free-text. A short, conversational, LLM-mediated survey that asks 3–4 adaptive questions and follows up on what the respondent actually said gets 4–6x the response rate and usable qualitative answers in every language the workforce speaks.

This is the channel that lets a bottom-up AI adoption program scale to 9,000 employees without 9,000 interviews. Done right, it produces a heatmap of friction by function, by geography, by tenure, by language. Done badly, it produces the same survey fatigue as every other corporate questionnaire.

Channel 3: Passive Telemetry (calendars, docs, tickets, systems already in use)

The third channel asks no questions of anyone. It reads the signals already being generated by the work itself: calendar density (where are the recurring 30-minute meetings that could be a Slack message?), document patterns (which templates get copy-pasted 200 times a month?), ticket clustering (which support categories spike every Monday?), tool usage (which licensed seats are dormant?). Most enterprises already have this data in Workday, Jira, ServiceNow, Google Workspace, Microsoft 365. Most do not look at it as an AI-opportunity signal.

Telemetry corrects for the bias in the first two channels: people tell you what they remember, not what they actually do. Calendar data tells you what they actually do.

The Three Channels Compared

Channel	Format	Who it reaches	Best signal	Output
Deep interviews	60–90 min, semi-structured, 1:1	30–50 workflow owners across layers	Causal narrative, exception detail	Opportunity hypotheses with workflow detail
AI-driven async surveys	5–7 min, conversational, multilingual	Full workforce (hundreds to thousands)	Distribution of friction, anomaly detection	Heatmaps by function / geography / language
Passive telemetry	No questions; reads existing systems	Everyone (whoever generates the data)	What people actually do, not what they say	Time-spend maps, queue clustering, dormant-tool flags

No single channel is sufficient. Interviews alone are biased toward articulate people. Surveys alone are biased toward whoever felt like answering. Telemetry alone is biased toward what your systems already measure. The discipline of bottom-up AI adoption is the cross-channel synthesis.

When Top-Down Is Actually Better

This post is selling a methodology, so it's worth being honest about where the methodology stops winning. There are real questions where top-down strategy work is the right shape.

• Portfolio decisions. "Should we acquire X, divest Y, enter market Z?" The answer lives in market data, competitor capability maps, capital allocation models, and a small number of executive judgements. Interviewing 500 employees won't help; their workflows are irrelevant to the question.
• Capital structure and financing strategy. Whether to refinance the term loan or do a follow-on equity raise is not a workforce question. Top-down wins.
• Regulatory positioning. "How should we respond to the new EU AI Act regime?" Lives at the General Counsel / Chief Compliance Officer level, not in the workflow detail.
• Brand and category positioning. Pricing, positioning, segmentation — these benefit from outside perspective, not internal workflow telemetry.

Top-down strategy houses are good at these things because they have built the muscle, the benchmarks, and the relationships over decades. If your question fits that shape, hire McKinsey. If your question is "where is AI ROI actually hiding in our 5,000-person organisation?", that is not the same shape, and bottom-up AI adoption is the better instrument for it.

Top-Down vs Bottom-Up: Side-by-Side Comparison

Here is the comparison we use ourselves when scoping an engagement against a parallel Tier-1 proposal. The framing is deliberately concrete — "200-page deck" is shorthand for the standard McKinsey / BCG / Bain shape, and "three-channel discovery" is shorthand for the bottom-up AI adoption approach described above.

Dimension	Top-Down (Tier-1 strategy house)	Bottom-Up AI Adoption (three-channel)
Who gets interviewed	30–50 executives, mostly C-suite and direct reports	30–50 deep interviews plus full-workforce async coverage plus telemetry
Output format	200-page deck + executive summary	Persistent intelligence layer (interactive atlas, scored workflows, source attribution)
Deliverable lifespan	~6 months before it ages out	Living artefact; updated as the org changes
Typical cost (enterprise)	$1.5M–$5M+ for a 3–6 month engagement	$200K–$800K for an equivalent 8–12 week program
Time to first usable insight	8–12 weeks (steering committee, then findings)	2–3 weeks (live heatmaps from async + telemetry)
Who can use the deliverable	Board + ExCo. Workforce never sees it.	Every operator, scored by workflow they own
Update mechanism	New engagement, new fee	Refresh cadence built into the artefact
Vendor-replacement insight depth	"Adopt copilots in 6 functions"	"Workflow #47 in shared services Lyon has 14-min/invoice triage cost — 70% reducible"

The honest summary: top-down produces a credible narrative for the board. Bottom-up produces a roadmap operators can execute. If you need both — and most enterprises do — the cheapest way to get there is to run bottom-up first and let the narrative fall out of the evidence, rather than the other way around.

Case Study: IBM's Bottom-Up AI Transformation

IBM's internal AI program is the best public example of bottom-up AI adoption at scale, even though IBM doesn't market it that way. Reframed against the framework above, IBM's approach is process-by-process discovery executed across the workforce, not a top-down "let's deploy watsonx everywhere" mandate.

The structural lesson is in how IBM got there. They didn't write a top-down deck called "The Future of HR at IBM" and then deploy a monolithic system. They identified 80+ individual HR processes — benefits enrollment, leave requests, compensation inquiries, manager actions, onboarding, offboarding — and treated each as a separate discovery, automation, and measurement problem. Each one was a workflow surfaced through usage data and front-line input, not assumed from an org chart.

That is the bottom-up AI adoption pattern: find the workflow, score it, automate the right slice, measure it, move to the next one. Multiplied 80 times in HR alone, then repeated across finance, IT, sales, and operations, you get IBM's $4.5B number. A top-down assessment would have produced a 200-page deck recommending "AI in HR" as a theme, signed off, and quietly aged out of relevance. The bottom-up pattern produced an operating system. For the full breakdown, see our IBM AI transformation case study.

The Deliverable: Persistent Intelligence Layer vs 200-Page Deck

Even when a top-down assessment surfaces the right insights, the deliverable shape sabotages adoption. A 200-page PDF sitting on a SharePoint:

• Goes stale in 4–6 months as the org reorgs, hires, ships product, changes systems.
• Can't be filtered or queried — you can't ask "show me opportunities in Lyon shared services with <3-month payback".
• Loses source attribution — the recommendation says "consolidate vendor onboarding" but you can't trace it back to the 12 interviews that produced it.
• Lives at the board level, never reaches the operators who could act on it.

A bottom-up AI adoption program produces a different kind of artefact — a persistent intelligence layer. Concretely, that means:

• Heatmaps of friction and AI-opportunity scoring across the org, drillable by function, geography, business unit, and workflow.
• A workflow knowledge graph where each opportunity is linked to the interviews, survey responses, and telemetry that produced it.
• Source attribution on every claim — click into "shared services invoice triage" and see the seven evidence points behind the score.
• Refreshable cadence — the layer updates when you re-run a survey or pull a new telemetry snapshot, rather than requiring a new engagement.

This is the part of the comparison that does not show up in a side-by-side pricing table but ends up dominating ROI. A deck depreciates the moment it ships. A persistent intelligence layer compounds.

How to Run Bottom-Up AI Discovery at Your Organization

If you want to attempt this internally before retaining outside help, the five-step playbook below is what we'd hand to a Group Strategy or Chief of Staff team. None of it is proprietary — the moat is in execution, not the framework.

SUPALABS First-Party Data

FAQ

What is bottom-up AI adoption?

Bottom-up AI adoption is a structured enterprise discovery method that surfaces AI efficiency opportunities from every layer of an organisation — not just the executive cohort that a traditional top-down assessment interviews. It combines deep interviews with 30–50 workflow owners, AI-driven asynchronous surveys across the full workforce, and passive telemetry from systems already in use (calendars, tickets, document patterns). The deliverable is a persistent intelligence layer rather than a 200-page deck, and the opportunity coverage is materially deeper because it captures workflow friction that lives one or two layers below executive visibility.

How is bottom-up AI different from top-down strategy work?

Top-down AI assessments interview 30–50 executives, triangulate against an industry benchmark, and produce a prioritised use-case deck. They are good at narrative, board-credibility, and portfolio decisions. Bottom-up AI adoption interviews a similar number of workflow owners (not the same as executives), then layers async surveys and telemetry to reach the full workforce. The deliverable is operational rather than strategic: a queryable atlas of workflows scored for AI efficiency, with source attribution, that operators can actually execute against. Top-down owns "should we", bottom-up owns "where exactly and how".

When does top-down AI assessment beat bottom-up?

Top-down still wins for questions that don't depend on workforce-level operational detail: portfolio decisions ("should we acquire X"), capital structure, regulatory positioning, brand and category strategy, and any question where the answer lives in market data and executive judgement rather than internal workflow telemetry. If the question is "where is AI ROI actually hiding in our 5,000-person organisation", that's a bottom-up problem. If the question is "should we exit market Y", that's a top-down problem. Most enterprises need both at different times, for different decisions.

What does a three-channel discovery program cost vs a McKinsey AI assessment?

A traditional Tier-1 AI assessment (McKinsey, BCG, Bain shape) for an enterprise of 1,000–5,000 employees typically prices at $1.5M–$5M+ for a 3–6 month engagement. A bottom-up AI adoption program with comparable scope — deep interviews plus full-workforce async coverage plus telemetry integration — typically prices at $200K–$800K for an 8–12 week program. The cost differential is structural: bottom-up replaces partner-level interview time with AI-mediated async coverage for the breadth layer, and produces a persistent artefact that doesn't require a new engagement to refresh.

How many employees do you actually need to reach for "bottom-up" to mean something?

The answer is not "all of them". A well-designed three-channel program reaches deeply with interviews (30–50 workflow owners), broadly with async surveys (target 40–60% response across the full workforce), and ambiently with telemetry (everyone whose work generates a digital signal). The reach metric that matters is not raw headcount; it is coverage of the workflows that drive cost and revenue. A 5,000-person enterprise where you've reached the 200 workflow owners that touch 80% of operating cost is meaningfully bottom-up. A 500-person enterprise where you've only spoken to the ExCo is not.

Can we run bottom-up AI discovery internally instead of hiring an external partner?

Yes, with caveats. The framework is not proprietary — the five-step playbook above is genuinely the playbook. The hard parts to internalise are: (1) the async-survey tooling and prompt design that gets response rates above 40% instead of 8%, (2) the cross-channel synthesis discipline that resists confirmation bias, and (3) the workforce credibility to surface friction honestly — people tell external interviewers things they will not tell their boss's boss. Most enterprises that try internally produce a useful first pass and then bring in an external partner for the second wave to correct for these gaps.

Sources & References