1. Answering tenders without starting from zero
The problem
Large companies and public bodies often buy software through formal tenders, also called RFPs (requests for proposal). One large tender we worked on in 2026 had 11 parallel workstreams, including data protection, technical and AI questionnaires totalling around 300 rows, functional requirements, a contract to review and pricing. Each workstream pulled in a different team.
The frustrating part was repetition. Buyers ask many of the same questions, yet every tender felt like a fresh start, and what the team learned on one bid rarely made it to the next.
What we built
Every tender now gets one ticket, and that ticket is the single place for documents, decisions, drafts and reviews. From there, an AI pipeline prepares first drafts for the team to work on:
- It reads the full tender pack and lists every requirement, deadline and potential deal-breaker, plus clarification questions for the buyer.
- It researches the buyer and recommends whether the bid is worth pursuing.
- It drafts the questionnaire answers, the proposal, the cover letter and a review of the contract terms.
- It runs a final check on the finished set for completeness, consistency and claims nobody can back up.
The AI answers from an approved fact sheet, where every figure and claim has a source and an owner.
What stays human
The AI prepares, and people do everything that follows. Sales and experts in security, legal and product edit the draft answers, review the whole bid and submit it. They also take every decision at four checkpoints: whether to bid at all, the bid strategy, sign-off on the reviewed answers, and submission. Nothing reaches a buyer unless a person sends it.
After each submission, the team adds its best answers about Kontent.ai to our internal answer library, once a person approves them. Each tender leaves the next one better prepared.
The result
Far less time goes into assembling documents and searching for answers given before, and that time goes back to the customer. The sales team spends it on what wins a tender: understanding the buyer's needs, shaping the best solution for them and building the relationship.
2. Supplier invoices: from retyping to reviewing
The problem
Supplier invoices arrive as PDFs, and no two suppliers lay them out the same way. Someone in Finance had to open each file, find the supplier, amounts, dates and reference numbers, and type them into the accounting system. The work was slow and tedious, and a typo in an amount is easy to miss.
What we built
Classic OCR turns a scanned page into text, and someone still has to make sense of it. Finance needed more: an invoice that arrives in their ticket queue and leaves as an entry ready to post. That is what happens now:
- AI reads each invoice the way a person would, telling the total from the tax and the reference number, whatever the layout.
- It links the invoice to the right supplier record in the accounting system.
- It combines the invoice data with the supplier's details into an import file the accounting system accepts.
- All of it runs inside the ticket queue Finance already uses.
What stays human
Our accounts are audited, so every amount we book has to be right, and someone has to answer for it. That responsibility stays with people. A person in Finance checks each prepared entry against the original invoice in the accounting system and posts it only when it is correct. If anything looks off, nothing is posted until it is fixed. AI does the reading and the typing, and Finance stays accountable for what goes into the books.
The result
Finance now reviews prepared entries instead of keying them in by hand. The time that went into retyping goes into checking, follow-up with suppliers and the analysis that needs an accountant's judgment. And because every entry reaches the accounting system in the same format, a wrong figure stands out instead of hiding in a typo.
3. Security reports: fair answers at a growing volume
The problem
Like many software companies, we run a vulnerability disclosure program. Independent security researchers report weaknesses they find in our product, and valid findings can earn a reward.
The challenge is volume and quality. Many reports describe behavior that is intentional and documented, repeat something already reported, or fall outside the program's scope. Since AI writing tools became widespread, the number of low-effort, AI-generated reports has grown sharply. Every report in scope still needs a careful, fair answer, and each one takes time from security experts.
What we built
An AI agent now does the first pass on every report:
- It checks the report against the program's scope and our public product documentation, to spot behavior that works as designed.
- It looks for duplicates of earlier reports.
- It proposes a severity rating, weighing the CVSS score, our internal vulnerability and incident ratings, and the severity tiers commonly used in bug bounty programs.
- It drafts the internal ticket and routes it to the team that owns the affected part of the product.
- It drafts a professional, factual reply to the researcher.
What stays human
Nothing the agent proposes is accepted until our security team has reviewed it. The team decides whether a finding is valid, confirms or changes the severity, and decides whether it earns a reward. Only then does the ticket go to engineers and the reply go to the researcher.
The result
Researchers get consistent, well-reasoned answers, and engineers receive tickets only for issues the security team has already validated, with the context they need to fix them.
What the three projects have in common
Our advice for anyone starting out:
- Pick a process people dislike, or one that is hard to scale with the team you have. Repetitive work with clear inputs is the best starting point.
- Give the AI one source of truth. When a fact changes, change it in one place.
- Keep decisions with people. Design the checkpoints first: who approves, and at which step.
- Work where the work already happens. All three automations run from tickets the teams already used.
- Close the loop. Approved corrections flow back in, so every round starts from a better place.
The same ideas, applied to content
Everything above happened outside marketing, but the pattern carries straight into content work: structured inputs, approved sources, AI doing the mechanical steps and people approving what goes live. That's the idea behind Kontent.ai, the world's first agentic CMS.