The mandate: find what’s slowing us down
My employer, a Spanish IT consultancy in its startup years, was growing fast enough that its lead-to-live process had never been written down. Sales had one picture of it, delivery had another, and nobody had the whole one. The brief I got was refreshingly open: find the operational bottlenecks and propose what to do about them. No predefined scope, no method handed down. That freedom cut both ways, because whatever I found would be about the people I work with every day.
Research my colleagues could trust
Interviewing colleagues about what’s broken is not like interviewing users. The findings can read as blame, the interviewer stays in the building afterwards, and honest answers only show up if people believe nothing will be traced back to them. So I designed for trust first: sixty-minute semi-structured interviews with question sets clustered by role, an explicit safe-space framing at the start, and a consent promise I could keep: every answer anonymised in every material, no stance ever attributable to a person, and all recordings deleted after transcription. They were.
The protocol itself carried a probing toolkit for getting past polite generalities: letting silence do the asking, click-by-click walkthroughs of real work, anchoring on the most recent concrete example instead of “usually”, and what-if scenarios to surface constraints people had stopped noticing. Every question also existed in Spanish, so each person could speak in whichever language they think in.
Research design · sixty-minute semi-structured protocol, EN + ES
An interview built so honesty is the easy option
Question sets clustered by role, a consent design stated out loud before the first question, and a probing toolkit for when the answers stay polite.
The hour, as scripted
- 00–03Intro & consent. The safe-space framing, then permission to record.
- 03–13Role scan. What the week actually contains, beyond the job title.
- 13–30Process tracer. The lead-to-live flow from this person’s seat, step by step.
- 30–39Lived experience. Where mistakes start, which tools help and which slow down.
- 39–45The magic wand. Fix one thing in your week: what and how? Buffer to the hour.
“I am not evaluating you or your performance in any way. I’m evaluating only the processes to make your life easier.”
The consent framing, read to every interviewee, with the promise that every answer would be anonymised and no name attached anywhere. Recordings stayed local and were deleted after transcription.
The probing toolkit
The silence trick
After a short answer: nod, wait three silent seconds. People fill the silence with the real, detailed answer.
The click-by-click probe
“You mentioned you send it over. Walk me through the exact steps. What exactly do you attach?”
The recent-example probe
“Tell me about the last time you did this. Did it go perfectly, or was there friction?” Recent beats “usually”.
The what-if probe
“What happens if you skip that step? Does the next person notice immediately, or does it surface later?”
Every question existed in Spanish too, so each person could think out loud in their stronger language.
Everyone, in two languages
I interviewed everyone in the company, leadership included, most of them within a single week. The interviews ran on Teams with automatic transcription, and every transcript then got corrected by ear, because bilingual recordings and auto-transcription are not friends. I anonymised the corrected transcripts myself before any analysis touched them.
For the analysis I worked with an LLM (Gemini, picked for its large context window) as a sparring partner: brainstorming candidate themes across the full interview set, then checking every candidate back against the transcripts before it earned a place. The human owns the claims; the model widens the search. The result was six themed findings, which condensed into three strategic pillars.
Making the invisible process visible
Findings about process need a picture of the process. I mapped the lead-to-live flow as a swimlane diagram built from what people actually described, not from how anyone assumed it worked, and modelled decision-making as a pair of DACI matrices: who currently drives, approves, contributes to and is informed about each decision, next to an ideal-state version of the same matrix. The gaps between the two maps did the arguing for me. A stalled deal or an ownerless decision stopped being somebody’s fault and became a visible property of the system, which is something a team can fix together.
Process mapping · the lead-to-live flow and its decision rights
The maps that did the arguing
The swimlane shows the form the argument took: where the flow crossed team lines, and how friction clustered at the handoffs. Roles generalised and assignments illustrative; the real maps carried the specifics, and those stayed internal.
Decision rights, current state
| Decision | Leadership | Sales | Delivery | Ops |
|---|---|---|---|---|
| Qualify the lead | D A | I | ||
| Price the work | D A | C | I | I |
| Start the build | D A | C | I | |
| Approve the result | I | D A | ||
| Update the client | no clear owner | |||
Decision rights, ideal state
| Decision | Leadership | Sales | Delivery | Ops |
|---|---|---|---|---|
| Qualify the lead | A | D | C | I |
| Price the work | A | I | C | D |
| Start the build | A | I | D | C |
| Approve the result | I | C | A | D |
| Update the client | I | D | C | A |
D drives, A approves, C contributes, I is informed. The pattern the real pair exposed: functions wearing two letters at once (outlined) and decisions with no owner at all, against an ideal of one letter per role per step. Steps and assignments here are illustrative, not the real matrix.
From findings to choices
Twenty-two proposals came out of the analysis, and no small company can absorb twenty-two changes at once. So I split them deliberately: eleven quick wins that were cheap and implementable immediately, and eleven strategic recommendations that carried structural change or cost. The written diagnosis ran to twenty-two pages for the record; the room got a condensed co-validation deck focused on the three friction points that mattered most. Same evidence, two altitudes, because a document that protects the reasoning and a presentation that provokes a discussion are different tools.
Thematic analysis · six themes, three pillars
Six themes, condensed into three strategic pillars
Every finding earned its place by surviving a check back against the transcripts. Wording kept at theme level; the pillars gave the solutions their shape.
Theme 01
The handoff gap
Context between sales and delivery moved as memory and unstructured notes, not as a defined handoff.
Theme 02
Ambiguous vocabulary
The same words meant different things to different people, so a deal’s true state depended on who you asked.
Pillar 1
Synchronise sales and operations
Delivery never starts a project blind, and sales never promises a date without a capacity check.
Theme 03
Hero mode
Decision authority concentrated on single individuals, so everything queued behind the same few people.
Theme 04
Strategic gaps
Lead qualification, the public face and the security posture needed more deliberate attention than growth had allowed.
Pillar 2
Unblock leadership and strategy
Fewer decisions queueing on the same desks, so strategy gets its time back.
Theme 05
Quality on discipline alone
Delivery pace had outrun quality structures; good outcomes leaned on individual care instead of process.
Theme 06
Scattered knowledge
Information lived in side channels and silos, so finding the current truth took detective work.
Pillar 3
Stabilise delivery and quality
Professionalise the lifecycle, so quality stops depending on heroics.
Primary mapping shown; several themes fed more than one pillar. Eleven quick wins and eleven strategic recommendations hung off these three pillars.
A workshop built to force decisions
Presentations don’t change companies; decisions do. So the follow-up workshop framed every single proposal as a forced choice: let’s do this, let’s skip this, or maybe later. Nodding was not an available outcome.
I also used humor deliberately. The most sensitive finding got introduced by an AI-generated comic strip that let the room laugh at the situation before discussing it, and the workshop closed with the transcription engine’s funniest mishearings from the interview recordings. Both served the same purpose: keeping a conversation about a company’s weaknesses from feeling like a trial.
The outcome: all eleven quick wins accepted. A proposed new hire deferred on cost. Sprint-based planning consciously skipped in favour of simple Kanban. The rejections were as valuable as the approvals, because a “no” decided in the open beats a “yes” nobody acts on.
Facilitation · every proposal left the room with a decision
The “do · skip · later” mechanic
Twenty-two proposals, three possible verdicts each, and no fourth option called nodding. A sample of how the cards landed:
Let’s do this
Weekly traffic-light status
Delivery posts red, yellow or green for every active project, every week.
A shared vocabulary
One page defining the terms that kept causing misunderstandings, in both languages.
A dedicated tech-debt day
Maintenance gets a scheduled slot; if it’s not scheduled, it won’t happen.
All eleven quick wins landed in this column.
Let’s skip this
Sprint-based planning
Too much ceremony for the team’s size; simple Kanban chosen instead, in the open.
A “no” decided together beats a “yes” nobody acts on.
Maybe later
A new specialist hire
The right idea at the wrong price: deferred on cost.
Deferred in the open, with the reason on the table.
What actually changed
The workshop’s decisions became an action plan with a named owner for every item and a five-phase rollout, launched in a kick-off with an agreed check-in cadence. Some of the first deliverables I built myself, including a bilingual consensus glossary of the terms that caused the most misunderstandings between sales and delivery, and an operational mini-dictionary to back it. Its sharpest tool is a mandatory clarification rule for the three most dangerous words in the building: “confirmed” now always answers confirmed what, “new client” always answers by which criterion, and “signed” always answers signed by whom.
What stuck, and I can vouch for these because I watched them become routine: a weekly traffic-light status practice, the shared vocabulary, a recurring marketing check, a cybersecurity shift, a tooling shift, a dedicated tech-debt day, internal workshops, and protected deep-work hours. One more action item deserves its own mention: the decision to rebuild the company’s website, which became my next project and the flagship case study on this site. The diagnosis also lived on as the empirical foundation of my M.Sc. thesis.
After those first actions went live, I moved to that website full time and a colleague took over the consultation’s implementation, extending it with ideas of his own. I can vouch for the practices above; the rest of the roadmap left my hands, and I honestly don’t know which of the later items landed.
Implemented deliverable · bilingual consensus glossary + mini-dictionary
A dictionary for the words that cost money
Operational definitions for the terms sales and delivery kept understanding differently, in Spanish and English, backed by one blunt rule set.
Glossary excerpt
Lead
Unqualified initial contact: a person and a source exist, but not yet evidence of fit, need and a buying process.
Prospect
Qualified lead: fits the profile and shows intent, with an accessible stakeholder. “Why now” must be written down.
Opportunity
Active negotiation tied to a deal record: stage, value, close date and an agreed next step, or it isn’t one.
Potential
A possibility, not a commitment. Without a range and a probability it’s wishful thinking disguised as a status.
The three most dangerous words
“Confirmed”
always answers: confirmed what?
Discovery done, proposal sent, contract sent, signed by one party, signed by both, first payment in. Pick one.
“New client”
always answers: by which criterion?
A deal closed and signed by both parties, or a first payment received. One criterion, used every time.
“Signed”
always answers: signed by whom?
By the client, or by both parties, and since when. “Fully executed” means nothing until someone says which.
What I’d do differently
Listen first, structure second. Before the interviews I had written a full workflow framework for our project-management tooling, built on general agile principles and my own assumptions. It was in real use only briefly. When the tool’s subscription came up for renewal, we dropped both: the framework was too rigid for a company our size that in essence needed a simple Kanban, and the interviews had said as much. We moved work management to GitHub Projects instead, where the developers already lived. The framework wasn’t wasted, it became my clearest evidence that a process has to be fitted to the team you actually have, not to the team the textbooks assume.
Being embedded is access and bias at once. Researching your own employer gives you context no external consultant gets, and opinions about every finding before the data is in. I managed the bias by grounding every claim in the transcripts and letting the maps carry the argument. The consent design mattered more than any single finding: colleagues talk to you again next quarter only if the promises held. They did.
Hand over deliberately, not by momentum. My focus moved to the website at full speed and the roadmap went with the handover. It worked, my colleague ran with it, but I never built in a check-back, so today I can’t account for the second half of the roadmap’s outcomes. A consultant’s last deliverable should be a way to know what happened after they left the room.