AI, No-Code or Custom Software?

Start with the tool you already pay for. Build only what it can't do.

Start With the Workflow

Three facts decide the tool: what it can reach, who checks the result, and who keeps it running.

Answer those first. Most of the decision follows.

Which Workflow Should You Automate First?

Choose a job whose result your team can check. A bigger frustration can still be the wrong first build if nobody agrees what should happen.

Consider a fictional distributor weighing two jobs:

A weekly operations report. The team already agrees on the reporting period, source files and what each total means. The operations manager reviews the report against those files and checks that every expected source is present. That gives the first automation a result someone can accept or reject.

Refund decisions. Sales and operations disagree about which returns qualify and who can approve an exception. Automating that decision would encode an unresolved policy. Settle the eligibility rules and approval authority before building it. Routing refund requests to a person can still be useful while those decisions remain manual.

Under those assumptions, start with the report. First test whether an existing reporting feature can produce the agreed output and expose missing sources. If it can, use it. Consider no-code only if its connectors, checks and maintenance owner cover the job. Consider custom work only for a specific unmet requirement, such as applying agreed mappings between source fields when the available tools cannot do it.

Try the synthetic weekly-report example to inspect missing-source and inconsistent-data checks. For your own candidates, write down the output, its reviewer and the check they will make. For the job you defer, name the decision that must be settled before you revisit it.

Three Ways to Automate

Smallest Tool First

Each fits a different kind of workflow, and one business can use all three.

A Built-In Feature

Your CRM, help desk or accounting tool may already route, summarize or draft. Test it on real cases, and check it respects your permissions and review steps. It fits when the work stays inside one product.

A No-Code Workflow

Tools like Zapier, Make or Airtable connect systems quickly and are easy to change. They fit when the available connectors cover each step, and a named person maintains the workflow.

A Custom Integration

Software built around your systems. It fits when a specific requirement goes beyond what the available tools support: custom logic, the access a step needs, or how approvals are recorded.

Where AI Fits

AI

A model is good at messy input.

Reading emails, PDFs and scans. Sorting requests by what they ask for. Pulling fields from documents that never share a layout. Drafting a reply or summary for someone to edit. Finding answers in your own documents.

It is the wrong tool for anything that must be exact every time: prices, totals, approval limits, permissions. Those belong in rules.

A chat assistant like ChatGPT helps a person draft and research. A model inside a workflow is different. It needs defined inputs, a check on every output, and a person for the uncertain cases. See how one runs, or try an interactive purchase-order example.

Does AI Leave Less Work?

Fictional comparison; all figures are invented. A distributor wants less staff work on the same 20 quote requests: 16 routine and 4 exceptions. The agreed output is a draft with correct items, quantities and sheet prices, checked by the salesperson. Set this quality floor first; the implementation acceptance example shows how.

Both methods produce the same 18 accepted drafts and leave 2 unresolved requests on hold. Current handling completes all 18 manually, with no failed attempts. Assisted processing fails on 2 requests; manual recovery completes them within the 18 accepted. Count every attempted case.

Staff effort in minutes for this batch, current → assisted:

  • Preparation: 30 → 10.
  • Checking: 30 → 40.
  • Corrections: 10 → 20.
  • Exception recovery and manual completion: 20 → 25.
  • Share of recurring operating work: 10 → 15.

Total: 100 → 110 minutes. Faster preparation leaves more checking and recovery work. Record waiting time, one-time setup effort and recurring cash costs separately; these are unknown here, not zero. Any staff time freed is capacity, not automatic payroll savings.

Decision: change the approach. For this staff-effort goal, keep current handling. Test whether a built-in feature or no-code change reduces the checking burden before considering custom work.

Copy this blank comparison for your own cases:

  • Same case set, mix and period [ ]; agreed output and acceptance check [ ].
  • For each method: attempted [ ]; accepted [ ]; failed [ ]; completed manually [ ]; still unresolved [ ].
  • Current → assisted staff minutes: preparation [ ] → [ ]; checking [ ] → [ ]; corrections [ ] → [ ]; recovery/manual completion [ ] → [ ]; recurring operating work for this period [ ] → [ ]; total [ ] → [ ].
  • Separately for each method: waiting [ ]; setup effort [ ]; recurring cash [ ]; unknowns [ ]. Continue, change or defer [ ]; reason and next check [ ]. If the quality floor fails, resolve that before using time to justify continuing.

When No-Code Is Enough

No-Code

Often, at first.

A no-code workflow is cheap to try and quick to change. For many internal processes it is the right answer, and it can stay that way.

No-code can include review and error handling too. Zapier can pause a workflow for human review, and Make can route failed steps to error handlers.

The signs you are outgrowing one: failures nobody notices, logic spread across workflows only one person understands, costs that climb with volume, and a requirement the platform can't meet.

Before You Build

Four Questions That Decide It

What can the tools reach?

Check what each system's connectors can read and write, the permissions they need, and what data you can export. The number of systems alone doesn't decide it.

Who checks the result?

If a mistake reaches a customer, an invoice or a record, the workflow needs a review step, an exception queue and a log of who approved what. The costlier the error, the more belongs in rules and review. Check which tool supports each.

Who keeps it running?

Every option needs an owner who notices when a connector breaks, a field is renamed or a model's output drifts. A process that still changes every month is often best kept simple until it settles.

What will the vendor allow?

Check rate limits, usage-based pricing, data retention, and where an AI provider processes your data, before the workflow depends on it. Custom code can't reach data a system won't expose.

When It Needs Custom Work

If the answer is a custom integration, that is our AI workflow automation work: the model, the rules, the review queue and the connection into your systems. It runs through Leveraged Velocity, with the systems in scope agreed before work starts.

If the answer is a feature you already pay for, start there. Example first moves show what a first step looks like for orders, quotes, reports and support.

Who builds it is a separate decision: a fixed-price project, an in-house team, or an ongoing one. Compare the five ways to handle the work.

Not Sure Which One You Need?

Describe the workflow and the systems it touches. You'll get a first move to review: what to automate, and what stays with your team.