Who this playbook is for — and when to use it
This playbook is for owners, operations directors and sales leaders in mid-sized companies — especially where the team is small and processes run on people's memory, email and Excel. The best moment to start: when business growth begins to require additional headcount for administrative work rather than for client relationships. That's when automation stops being an "interesting project" and becomes an alternative to a growing cost structure.
The worst moment: when the team is in the middle of an operational crisis, or when the organization is in the middle of changing its core systems (CRM, ERP, accounting). First stabilize the foundations, then automate the layers above them.
Where we most often start
These five areas come up most often in our work with mid-sized companies — because they combine high value, available data and risk that can be managed.
Automation 1 — Quotes and tenders (RFP)
Who it's for: Manufacturers and suppliers with an extensive product catalog who respond to specifications from design offices, general contractors or public buyers. The sales and project team spends hours reading tenders and manually matching specifications to products.
How it works: The system supports three steps. First it analyzes the tender and extracts the technical requirements, quantities and evaluation criteria. Then it matches products from the catalog — with card indexing, metadata enrichment and RAG search that understands context, not just keywords. Finally it generates a draft quote: a ranking of matches that can be corrected by hand, a PDF export and suggested strategic questions for the client.
What it changes: Sales stops playing defense. Specifications are read faster, quotes have a consistent standard, and salespeople can return to relationships and strategy instead of hunting for matches in the catalog. The same team handles more inquiries without raising headcount — and the company gains room to start shaping demand earlier (e.g. with materials for design offices) instead of only reacting to other parties' specifications.
Automation 2 — Email and daily summaries
Who it's for: Companies where the inbox is the de facto operating system — salespeople, project managers and support work in email, and triage, classification and replies eat up hours a day. We see this most often in B2B services, in projects and in key-account teams.
How it works: The agent reads the incoming inbox (Microsoft 365 or Google Workspace), classifies messages by project, client or case type, flags those that need an urgent response, and generates a daily summary for a person or a team. Important matters keep their full context; recurring ones get a draft reply for a human to approve. All within the bounds of permissions — the agent doesn't reply on a person's behalf without their consent.
What it changes: The first hour of the day stops being inbox triage. The team sees what genuinely needs a decision and what can wait. A key client doesn't get lost in newsletter noise. Sales and support reclaim time for work that creates value — and the owner gains a view of the team's working rhythm that previously couldn't be measured.
Automation 3 — CRM and ERP — leads, opportunities, pipeline
Who it's for: Companies where leads live in email, in Excel and in salespeople's heads, and sales has no single queue. This is often accompanied by a product configurator, a website form or a marketplace entry point — which generate leads but don't feed them into any process.
How it works: A workflow in a system like ERPNext (or in the existing CRM) has three layers. The first is intake — an HTTP API or a form integration that creates a lead, deduplicates it and assigns an owner. The second is lead → opportunity automation, with a pipeline workspace, a Kanban view and stage gates (interest, qualification, quote, deposit). The third is the permission model: who sees what, who can edit, what the escalation path looks like.
What it changes: Every lead lands in one place, with one owner and a clear stage. Sales sees the pipeline in real time; management gets predictability where there used to be administrative work and intuition. The company stops losing some opportunities just because someone didn't copy data from an email into the system.
Automation 4 — Documents and product metadata
Who it's for: E-commerce and companies with an extensive product catalog or a large media library (product photos, spec sheets, technical documentation, certificates). Working with metadata consumes manual hours, and inconsistent file names and descriptions break internal search engines and SEO.
How it works: The system tidies the catalog on two levels. At the image level — automatic renaming, classification by content, tagging and linking to SKUs. At the product-card level — filling in descriptions, extracting attributes from technical documentation, normalizing units and translations. All within an approval loop — a human sees the proposed change and accepts or corrects it.
What it changes: A product card that previously required half an hour of manual work becomes a proposal to approve in a few minutes. Internal search and the store start finding what the customer types, not only what exactly matches the SKU name. The team that used to maintain the catalog can move on to developing the product rather than maintaining it.
Automation 5 — Knowledge systems and RAG
Who it's for: Companies where knowledge about products, projects, clients and processes is locked in people's heads and old files. Onboarding a new person takes too long, RFPs require digging through the archive every time, and "what did we do back then" is a question no one can answer without a conversation.
How it works: The knowledge system has three elements. The first is capture — an interview agent runs a conversation across domains (company history, services, competencies, projects, compliance), generating structured Markdown chunks with metadata and a confidence rating. The second is retrieval — RAG with filters by domain, topic, type, confidence and tags. The third is provenance — every answer comes back with a link to the original chunk, a date and who added it.
What it changes: The company's memory stops being the private property of a few people. Onboarding shortens, RFPs become repeatable, and the answer to "what did we do back then" reaches anyone with the right permissions. Knowledge doesn't walk out the door with people, and the model doesn't "hallucinate" company history — because it always answers from a specific, verifiable source.
Three criteria that order the sequence
There's no single "best" automation — there's a first, a second and a third. We order the sequence by three criteria:
Criterion 1 — Value. Where is the return from automation greatest over a year? What counts is the team's time and the quality of decisions, not the mere fact that "AI runs there." Most often the winner is the workflow that today consumes the most manual hours and recurs daily or weekly.
Criterion 2 — Data. Do we already have the data the AI is meant to operate on, or do we have to build it first? Where data exists in a sensible form (product cards, mailbox, CRM, project archive), a pilot starts faster. Where you have to begin by civilizing the data, the project starts with less spectacular but more important work.
Criterion 3 — Risk. What are the consequences of the agent making a mistake? A workflow where an error costs a minute of a person's time can be piloted quickly. A workflow where an error costs a client, a transaction or regulatory compliance requires sharper boundaries, an audit trail and a human in the loop — and shouldn't be the first pilot.
The best first project is one that combines high value, available data and low risk at once. If there's no such project — we start from the highest value and cautious boundaries.
Three most common blockers — and what to do about them
Blocker 1 — No process owner. Automating a workflow that no one in the company owns will always stall. First decide who in the organization is responsible for "this process" (sales, operations, the procurement director, the owner) — only then the project. Without an owner, the pilot has no one to approve it and no one to correct it.
Blocker 2 — Pilot too broad. All "transformations" fail for the same reason — the scope is too big to finish in a reasonable time and show value. The first pilot must have a narrow definition: one lead segment, one email category, one tender type, one knowledge domain. Expansion comes after the first wave of return, not before it.
Blocker 3 — No boundaries for the agent. "AI will do it for us" is the worst brief. The agent must have clear boundaries — what it may do on its own, what it may only propose for a human to approve, what it shouldn't touch at all. Without this, the pilot will generate concern from the team and the board before it generates value. Boundaries are part of the product, not an obstacle.
Who benefits most
SMB / mid-market
This playbook is written first and foremost for you. The fastest lever is the first concrete workflow, not an AI strategy or an AI department. Pick one area from the five above, match it against the criteria (value, data, risk), and start with a pilot — not a program.
Enterprise
For leaders of large organizations, this playbook serves as a filter: which areas make sense as first pilots in subsidiaries or operating units. Each of the five automations also scales to enterprise — with stronger governance, auditability and AI gateways.
Private Equity
For funds and portfolio operators, the five serve as a value-creation map for the SMB companies in the portfolio. Each of the five areas has a clear operational return and clear risk — they can be put into the value map and the portfolio playbook without guessing "what AI will change in this company."
One step you can take this week
Pick one workflow in the company that today consumes the most of the team's time and recurs daily or weekly. Write down in three sentences: what problem it solves, what data it needs, and what would happen if the agent got it wrong. Those are the three answers we start every pilot conversation with — and they immediately give you the material for a sensible brief.