Steply
 Book a briefing
Cases

A result is not what the vendor says. It is what the indicator shows.

Every case below has the same structure: the bottleneck that existed, what went into production and the indicator that started being tracked in the dashboard. Where the starting number is known, it is right there. A percentage with no baseline is an ad.

How to read a case, including the ones that are not ours.

A case page is the easiest thing for a tech company to dress up. These three questions separate proof from advertising, and they work on any vendor you are evaluating.

One big number, alone, in a huge font.
Off what baseline? "−80%" on a process that took 10 minutes saves 8 minutes. Without the starting point, the percentage states no size at all, which is exactly why it usually comes without one.
The technology stack, proudly listed.
And what changed in the operation? Stack is a builder's decision, not a buyer's benefit. What matters is which person stopped doing which task, and what they do instead now.
The result measured once, in the month it shipped.
Who measures it today? A gain that only exists in the closing deck evaporates in three months. If the indicator does not stay in the client's dashboard after the project, neither client nor vendor knows whether it held.

Six operations, six bottlenecks.

B2B SaaS

Customer onboarding that runs itself

What was jammed

New-client onboarding was manual, took 14 days and lost 38% of people before first use. Every new account depended on someone remembering to chase the missing document.

What went live

An agent guides the client step by step, reads submitted documents, validates what can be validated on its own and opens an internal ticket only when a human is needed. Every onboarding step became an event in the dashboard.

What started being measured

−72%

time to first use

before
14 days
after
3.9 days
Composition
  • Sistema web
  • Agente com leitura de documento
  • Integração com o sistema interno
Fintech

Signup analysis with humans only on the doubtful cases

What was jammed

Signup analysis was manual, took 4 days per client and cost R$ 23 per analysis. It was a direct cap on growth: selling more meant hiring more analysts.

What went live

A pipeline reads the documents, classifies them and applies the business rules that used to live in the analysts' heads. Humans review only what the pipeline itself flags as ambiguous, and that rate is tracked in the dashboard.

What started being measured

−83%

cost per analysis

before
R$ 23.00
after
R$ 3.90
Composition
  • Leitura automática de documento
  • Regras de negócio versionadas
  • Integração com o cadastro existente
Healthcare

WhatsApp scheduling with no phone queue

What was jammed

Reception was permanently overloaded, with peaks of 17-minute phone waits. A patient who gives up scheduling shows up in no report: the cost was invisible.

What went live

An AI attendant on WhatsApp, wired into the real calendar: books, reschedules and answers common questions. Reception gets only the cases needing judgment, and the queue became a visible metric.

What started being measured

−65%

wait time at peak

before
17 min
after
6 min
Composition
  • WhatsApp Business
  • Agente conectado à agenda
  • Painel de fila e de desistência
E-commerce

A 120k-product catalog rewritten in bulk

What was jammed

120k products carrying the generic supplier description, identical to every competitor selling the same item. With no original text, search engines had no reason to prefer this store.

What went live

A batch process rewrites title, highlights and description for each product from its real attributes. Sampled human review on the highest-margin categories, with the correction rate tracked.

What started being measured

+28%

organic traffic conversion

change against the previous period average

Composition
  • Processamento em lote
  • Revisão por amostragem
  • Integração com a plataforma de e-commerce
Logistics

What the driver knew became a routing rule

What was jammed

The routing system ignored constraints only drivers knew: streets the big truck cannot enter, stores that only receive until 11am, gates that close for lunch. Every day the "optimal" route was fixed by hand.

What went live

The notes drivers were already writing become structured rules and feed back into the routing system daily, with no typing. Every rule created stays visible in a panel, with who originated it.

What started being measured

−18%

kilometres driven

change against the previous period average

Composition
  • Leitura de texto livre
  • Integração com o sistema de rotas
  • Painel de regras com autoria
Education

An assistant that helps without giving the answer

What was jammed

A student got one exercise wrong and dropped the course right there. Engagement and completion fell month after month, and the platform only found out in the next quarter's report.

What went live

An assistant that knows that course's material and answers in levels: the hint first, then the path, and only then the worked example. Drop-off per exercise started being measured in real time, per module.

What started being measured

+41%

module completion

change against the previous period average

Composition
  • Plataforma web
  • Assistente com o material do curso
  • Painel de engajamento por módulo

What the six have in common is not the technology.

Different industries, with problems that look nothing alike. What repeats is the rule: none of them started big, none of them ran outside the client's own house, and all of them ended up with a number tracked after delivery.

  • All started from one stage

    None of these started as a "complete platform". The first stage cleared a single bottleneck, with scope, date and price fixed. The rest only existed because the first one worked.

  • All run in the client's own house

    Environment, repository and keys in the buyer's name. In the cases involving sensitive documents, the AI model also runs inside the infrastructure: data does not leave the network to be processed.

  • All left a number in the dashboard

    Each case's indicator was not measured only at delivery: it stays in the client's dashboard, next to availability, errors and AI cost. That is what makes it possible to know whether the gain held after the excitement faded.

Which of these looks like yours?

It does not have to be the same industry. What repeats is the shape of the problem: someone repeating by hand a task the system should do, with nobody able to say what it costs per year.

Measure my bottleneckSee how scope gets locked