# Build an AI Native Business Redesign how your company works with AI. Learn where AI belongs in your company, what to automate or delegate, and how to redesign workflows around AI. ## Course format 8 modules, 10 lessons, and a final business blueprint. ## Course lessons Start here ## Build an AI-Native Business Learn where AI belongs in your company, what to automate or delegate, and how to redesign workflows around AI instead of bolting it onto the ones you already have. Course outcome Founders leave with one redesigned AI workflow, an AI opportunity map, and a practical 90-day implementation plan. What you will produce AI Opportunity Map Every recurring workflow in your business, scored by value and difficulty. Delegation Matrix A defensible line between human-led, AI-assisted, and AI-executed work. One redesigned workflow End to end: triggers, inputs, decisions, tools, approvals, outputs. 90-day roadmap Sequenced pilots, owners, metrics, and a governance model that ships with them. How to work through it 1 Read a module, then do its exercise before moving on. Each exercise produces one page of your blueprint. 2 Work on one real business, not a hypothetical one. Pick workflows you could change this quarter. 3 Modules 1–4 change how you see the work. Modules 5–8 make it operational. The final project assembles both halves. Before module 1 Write down the three parts of your business that feel most expensive in time and least differentiated in skill. You will test those guesses in module 2. Module 01 · 18 min ## What an AI-Native Business Actually Is The difference between a company that uses AI and a company built around it shows up in the work, not in the tool list. One buys licences. The other changes who does what, in which order, and who signs off. Key ideas • Using AI tools vs rebuilding work around AI • Assistants, automation, agents, and autonomous workflows • Why adding AI to broken processes produces limited value ### Four levels of AI work Most confusion about AI strategy is really confusion about which of these four you are talking about. They differ in who chooses the next step and who owns the outcome. Level 1 Assistant A human is in every step and owns the result. AI drafts, summarises, rewrites. Value scales with individual skill, so it rarely shows up in company metrics. Level 2 Automation Fixed rules on a fixed path. Predictable and auditable, brittle when inputs vary. Best for high-volume steps with one correct answer. Level 3 Agent Given a goal and tools, it chooses the steps. Handles variation, needs explicit boundaries, and produces work a human approves before it leaves the building. Level 4 Autonomous workflow Runs on triggers without a human in the loop, with monitoring, sampling, and a defined stop condition. Reserved for reversible, low-risk work. ### Why bolting AI onto a broken process disappoints A process that takes eleven days usually spends nine of them waiting: for an approval, a file, a reply, a meeting. Speeding up the two working days changes almost nothing. Worse, AI applied to a step nobody should be doing makes the useless step cheaper to keep. This is why the companies seeing real financial impact redesign the flow before they add the model. Do this now · 15 min List every place AI is already used in your company today. Label each one Level 1–4, and next to it write the business number it moved. Most lists come back as Level 1 with no numbers attached. That gap is the course. Module 02 · 25 min ## Map How Your Business Works You cannot redesign what you have not written down. This module turns the vague sense that "sales admin eats our week" into a list of named workflows with volumes, owners, and failure points. Key ideas • Identify recurring workflows across sales, marketing, support, operations, and finance • Find bottlenecks, repetitive work, delays, and knowledge dependencies • Create an AI Opportunity Map ### The workflow inventory Aim for ten to twenty rows, two to four per function. One line per workflow, in this shape. Workflow Trigger Runs / wk Hrs / run Where it stalls Inbound lead qualification Form submit 60 0.4 Research by hand Monthly close reconciliation Calendar 0.25 14 Waiting on receipts Tier-1 support replies Ticket 180 0.2 Answers live in one head Score each row: weekly hours = runs × hours , then note whether the pain is volume, delay, quality, or dependency. Four different pains, four different fixes. ### Four signatures worth hunting Repetition The same shape of work, many times a week, with small variations. Highest-value AI territory. Delay Work sitting in a queue. Fix the queue design first; AI cannot shorten a wait for a human decision. Knowledge dependency Only one person can do it. Writing the knowledge down is both the AI enabler and the risk fix. Rework Output that comes back for a second pass. Signals an unclear standard, which AI will faithfully reproduce. ### The AI Opportunity Map High value · Low effort Start here. Two workflows, maximum. High value · High effort Fund deliberately, in quarters, with an owner. Low value · Low effort Fine for training the team. Never the strategy. Low value · High effort Delete the workflow instead. Value = hours saved, revenue influenced, or cycle time removed. Effort = data readiness, integrations needed, and risk of being wrong. Do this now · 45 min Build the inventory with the person who actually does the work, not from memory. Then place every row on the map and circle the five in the top two quadrants. That circle is page one of your blueprint. Module 03 · 20 min ## Decide What to Augment, Automate, or Delegate Deciding what belongs to a human and what belongs to AI is a design decision, made once per workflow and written down. Teams that make it explicitly get further than teams that leave it to individual habit. Key ideas • What AI should assist with • What AI can execute independently • What must remain human-led • Evaluate work by judgment, risk, repeatability, and reversibility ### The four-question test Score each step of a workflow from 1 to 5. The answers, not opinions about AI, decide the mode. Judgment How much does a good answer depend on context that is not written down anywhere? High judgment stays human-led. Risk What is the cost of one bad output reaching a customer, a regulator, or a bank account? High risk keeps an approval gate. Repeatability Does it follow a recognisable pattern with a definable standard of "good"? High repeatability is where AI compounds. Reversibility If it goes wrong, can you undo it cheaply and quickly? Reversible work is the safest place to remove the human from the loop. ### Reading the scores Mode Score pattern Example Human-led High judgment, high risk, low reversibility Pricing exceptions, hiring and firing, partnership terms Augment High judgment, high repeatability in the inputs Deal reviews, hiring scorecards, board narrative drafts Delegate Low judgment, moderate risk, reversible, with review Lead research briefs, first-draft replies, competitor digests Automate Low judgment, low risk, high repeatability, fully reversible Data entry, routing, tagging, status updates, meeting notes ### Two failure modes Over-delegation is loud: a customer gets a wrong answer and everyone learns. Under-delegation is quiet: senior people keep doing work that scored low on all four questions, and nobody notices the cost. Review the matrix quarterly, because scores move as your standards get written down. Do this now · 30 min Take the five workflows you circled in module 2. Break each into steps and score every step on the four questions. Assign a mode per step, not per workflow. Most workflows come out mixed, and that mix is your Delegation Matrix. Module 04 · 30 min ## Redesign Workflows Around AI This is the module that produces the deliverable. You will take one workflow and rebuild it from the outcome backwards, rather than annotating the version you inherited. Key ideas • Start with the desired outcome, not the existing process • Define triggers, inputs, decisions, actions, tools, and outputs • Design exceptions, human handoffs, and approval points • Remove unnecessary steps instead of merely accelerating them ### The six-part workflow spec One page per workflow. If you cannot fill a field, you have found the work to do before any AI is involved. Trigger What starts this, precisely. An event, not a time of day, wherever possible. Inputs The data and context required, and where each piece lives. Name the system, not "our files". Decisions Every judgment made along the way, with the rule or standard that governs it. Unwritten rules become invented rules. Actions What actually changes in the world: a record written, a message sent, a payment scheduled. Tools The systems and permissions needed to take those actions, and the scope of each. Output The finished thing, in a defined format, with the quality bar written as a checklist. ### Worked example: inbound lead Before · 6 steps, 2 days Form lands in a shared inbox SDR checks it twice a day SDR googles the company for 15 min SDR writes notes into the CRM SDR drafts a reply from a template Manager spot-checks a few After · 3 steps, 4 minutes Submit triggers the workflow directly AI enriches, scores against the written ICP, writes the CRM record, drafts a reply with sources SDR approves or edits in one screen; anything below the score threshold is auto-nurtured Two steps were deleted, not accelerated: the inbox sweep and the manual note-taking. ### Design the edges before the happy path Workflows fail at their edges. Decide these three in writing: Exception What counts as "I do not know", and what happens then. A named queue, not silence. Handoff What a human receives at the handoff: the draft, the reasoning, the sources, and the specific question to answer. Approval Which actions require a signature, who can give it, and what happens if nobody does within the deadline. Do this now · 60 min Pick your highest-value workflow. Write the outcome in one sentence, then fill the six-part spec for the version you would build if the current process did not exist. Count the steps you deleted. If the number is zero, you have documented, not redesigned. Module 05 · 22 min ## Build the Human–AI Operating Model A redesigned workflow only holds if the org around it changes too: who is accountable, what a good week looks like, and what managers are now expected to be good at. Key ideas • Humans set direction, standards, and accountability • AI handles research, synthesis, execution, and coordination • Redesign roles, responsibilities, KPIs, and management practices • Teach teams to supervise AI rather than simply prompt it ### The split Humans own Direction and priorities The definition of good Relationships and negotiation Final accountability for every output Deciding what not to do AI carries Research and enrichment Synthesis and first drafts Repeatable execution at volume Coordination, follow-ups, status Monitoring and flagging exceptions ### Four skills to build deliberately Prompting is the smallest of them. Supervision is the job. Briefing Stating the outcome, constraints, audience, and quality bar without a conversation. Reviewing Sampling output against a checklist instead of reading everything, and knowing when to raise the sample rate. Escalating Recognising when to stop the workflow, and having the authority to do it without a meeting. Improving Turning each correction into a permanent change in the standard, the data, or the spec. ### KPIs move up a level Was measured as Becomes Emails sent, tickets closed Resolution quality and repeat-contact rate Drafts produced Approval rate on first pass Hours worked on a process Cycle time and cost per completed unit Individual output Workflow throughput and exception rate Do this now · 30 min Rewrite one role description for the workflow you redesigned. State what the person now owns, what they supervise, what they no longer do, and the two numbers they are measured on. Then name, by person, who is accountable when the AI gets it wrong. Module 06 · 24 min ## Give AI the Foundations It Needs A capable model with no access to your business is a stranger with good manners. Foundations are what turn it into a colleague: knowledge it can trust, tools it can use, limits it cannot cross, and a way to tell whether it is doing well. Key ideas • Business knowledge and trusted data • Tools, integrations, and system access • Permissions and access boundaries • Monitoring, testing, memory, and feedback loops ### Readiness audit Knowledge One source of truth per topic, with an owner and a review date. If two documents disagree, the AI will confidently pick one. Start with the five things it must never get wrong: pricing, ICP, positioning, policies, tone. Tools and integrations List the actions each workflow needs, then the system that performs each. Read access is cheap to grant and cheap to reverse; write access deserves a decision per action. Permissions Treat every AI worker as a named employee with least-privilege access, its own credentials, and a scope you can revoke in one place. No shared admin keys, no borrowed logins. Evaluation and monitoring Collect twenty real past cases with known-good answers and run them as your test set before and after every change. Log every action taken, sample outputs weekly, alert on volume, cost, and exception spikes. Memory and feedback Decide what should persist between runs: preferences, corrections, account history. Every human correction should update the standard, not just this one output. ### Sequence matters Do not wait for a perfect data estate. Fix the foundations for one workflow at a time: the knowledge that workflow reads, the tools it touches, the tests that cover it. This is how a company builds a real data and governance layer without a two-year programme first. Do this now · 40 min Score your redesigned workflow red, amber, or green on all five foundations. Fix the reddest one this week, and write your twenty-case test set before you build anything. Module 07 · 20 min ## Govern AI Without Blocking It Governance fails in two directions. Too little and one bad output becomes a customer incident. Too much and every idea waits in a committee. The fix is tiering: match the weight of the control to the consequence of being wrong. Key ideas • Risk levels and approval rules • Quality control and fact-checking • Privacy, security, and unintended actions • Escalation paths and clear human accountability ### Three risk tiers Tier 3 · low Internal, reversible Notes, summaries, internal drafts, tagging. No approval. Weekly sampling and logging only. Tier 2 · medium Customer-facing, recoverable Outbound replies, published content, CRM writes. One named approver, a checklist, and a full audit trail. Approval can be relaxed once first-pass accuracy holds above your threshold for a month. Tier 1 · high Money, contracts, personal data, legal Payments, pricing commitments, anything regulated. AI prepares, a human decides, always. Two-person sign-off above a threshold, and no autonomous execution. ### Controls that earn their cost Cite or abstain Factual claims arrive with a source, or the workflow says it does not know. This single rule removes most quality incidents. Scoped access Least privilege per workflow, no standing write access to systems it does not need, and a documented data-handling rule for anything personal. Action limits Caps on volume, spend, and recipients per run, so an error is small before anyone notices it. Escalation path A named person, a channel, and a stated response time. Plus one person who can switch a workflow off without asking permission. Accountability Every workflow has one human owner. "The AI did it" is not an answer a customer or a regulator accepts. Do this now · 25 min Assign a tier to each of your five workflows and write the approval rule for each in one sentence. Then name the owner and the person who can pull the plug. One page, and your governance model exists. Module 08 · 26 min ## Measure Value and Scale Pilots that cannot show a number get quietly defunded. Decide what you are measuring, capture the baseline before you change anything, and scale only what beats it. Key ideas • Measure revenue, cost, capacity, quality, and cycle time • Select one or two high-value workflows first • Pilot, evaluate, improve, and scale • Build a 30/60/90-day AI roadmap ### Five metric families Family Measure Baseline needed Revenue Pipeline created, win rate, expansion Last two quarters Cost Cost per completed unit, including tool and model spend Current cost per unit Capacity Units handled per person per week, hours redirected Two weeks of tracking Quality First-pass approval rate, error rate, rework, CSAT Sample of 20 past outputs Cycle time Trigger to finished output, including waiting time Ten recent runs Capacity freed is only value once it is redirected to something you can name. Write down where the hours go. ### The 30/60/90 roadmap Days 1–30 · prove One workflow live, baseline captured Test set of 20 cases written Owner named, tier assigned Weekly review of every exception Days 31–60 · harden Approval loosened where accuracy holds Second workflow started Knowledge sources consolidated Roles and KPIs updated Days 61–90 · scale Results reported against baseline Pattern reused in a second function Governance written into onboarding Next two workflows funded ### Stop rules Decide in advance what would make you kill a pilot: accuracy below a stated bar after two improvement cycles, cost per unit above the human cost, or an exception rate that consumes more supervision time than the work it replaced. Written stop rules are what make it safe to start. Do this now · 45 min Pick two metrics per family for your first workflow, record today's baseline, and fill the 30/60/90 grid with dated commitments and an owner per line. Final project ## AI-Native Business Blueprint Six artifacts, assembled from the exercises you have already done. Together they are a document you could hand to your leadership team on Monday and start executing. 01 AI Opportunity Map Every recurring workflow, placed by value and effort, with the five candidates circled. From module 2. 02 Human/AI Delegation Matrix Each step scored on judgment, risk, repeatability, and reversibility, with a mode assigned. From module 3. 03 One redesigned end-to-end workflow The six-part spec, plus exceptions, handoffs, and approvals, with deleted steps marked. From module 4. 04 Governance and approval model Risk tier, approval rule, escalation path, and named owner per workflow. From modules 6 and 7. 05 ROI scorecard Baseline and target for revenue, cost, capacity, quality, and cycle time, with stop rules. From module 8. 06 90-day implementation roadmap Dated commitments across prove, harden, and scale, each with an owner. From module 8. The quality bar ✓ A stranger could run your redesigned workflow from the spec alone. ✓ Every AI-executed step names the human who is accountable for it. ✓ At least one step was deleted, not accelerated. ✓ Every number in the scorecard has a baseline you measured, not estimated. ✓ Day 1 of the roadmap starts within two weeks. Reference ## Sources and further reading The course rests on three source foundations. Read them in this order if you want the underlying evidence. McKinsey, State of AI 2026 AI high performers redesign workflows instead of inserting AI into existing processes, combine efficiency with growth, and measure business impact. mckinsey.com — The State of AI: Global Survey 2026 Microsoft, 2026 Work Trend Index Effective AI users deliberately decide what belongs to humans versus AI, while humans retain judgment, quality control, direction, and responsibility. microsoft.com — Agents, human agency, and opportunity BCG, AI-First Enterprise Operations Transformation should begin with business outcomes, redesign processes end-to-end, and establish data, integrations, governance, evaluation, and monitoring from the start. bcg.com — Reinventing the Operating System of Work with AI