How to shortlist Upwork proposals for a chatbot or automation job

The short answer

When your chatbot or automation job gets twenty or thirty proposals, most of them will look alike: a friendly opening, a list of tools, a promise of quick delivery. The differences that matter are small and easy to miss.

Ignore where a proposal sits in the list and how polished it reads. Shortlist on four signals: it answers your screening questions about your job, it attaches work close to yours, it asks about failure, and its record fits the size of your project. Then ask your three shortlisted freelancers the same three questions and decide on the answers.

What to ignore

Position in the list. Upwork lets freelancers bid extra Connects to boost a proposal. The highest bidders take the top four slots and get a Boosted icon, according to Upwork’s help center. That is a bid for visibility. It says nothing about fit.

A polished cover letter. Writing a good proposal is a skill, and AI tools have made it cheap. A proposal that restates your job post fluently tells you the freelancer read it, not that they have built it.

Badges on their own. Badges carry real information, but each one answers a narrow question:

Signal	What it measures (Upwork help center, Oct 2026)	What it does not tell you
Job Success Score	Client feedback, how past contracts ended, higher-value projects and long-term relationships; the best of 6-, 12- and 24-month scores	Whether they have built your kind of chatbot or workflow
Top Rated	A Job Success Score of at least 90% for 13 of the last 16 weeks	How they handle a job your size
Rising Talent	A complete profile, recent activity, a verified identity and no negative client feedback	Depth of experience
Expert-Vetted	Passed Upwork’s invitation-only evaluation, including an interview	Fit to your systems
Identity check (blue tick)	The freelancer completed identity verification	Anything about the work

Star ratings. On Fiverr, 94% of the 317 automation reviews I read were 5 stars. Ratings that nearly everyone has cannot separate two good-looking freelancers on any platform. Feedback text is more useful than the score, when it describes a job like yours.

The four signals that matter

1. It answers your screening questions about your job. You can add screening questions to a job post. If yours ask about your workflow (“If the chatbot does not know the answer, what should it do?”), the answers are the most useful text in the proposal. A vague answer to a specific question is a clear signal.

2. It attaches work close to yours. The closest thing they have built to your project, shown briefly. My own decisions on sellers came from exactly this: the similar past work they attached in reply to my first message. As far as I remember, what was attached were mockup designs rather than live links, which is common, so treat samples as evidence of fit rather than proof that something ran in production. The portfolio guide covers how to read them.

3. It asks about failure. For a chatbot: what happens when it does not know, when a question is out of scope, when a person needs to take over. For an automation: what happens when a step fails or an input is missing. A proposal that asks one of these questions has thought past the demo.

4. The record fits the size. A freelancer whose history is mostly small, quick jobs may be exactly right for one workflow, and stretched on a months-long chatbot project with integrations. Read the feedback for jobs of similar size and length.

Two proposals, side by side

Say your job is a support chatbot for a small online shop. It should answer delivery and returns questions from your policy pages, and pass anything about a damaged order to a person. One of your screening questions asks: “What should the chatbot do when a customer asks something the policy pages do not cover?”

Proposal A sits first in the list with a Boosted icon. It opens with a warm paragraph about your business, lists six AI tools, promises delivery in three days and answers the screening question with “I will make sure it handles all questions perfectly.”

Proposal B sits further down. It is shorter. It answers the screening question with “it says it does not know, offers the contact form and logs the question so you can add it to the policy pages.” It attaches a short recording of a similar returns bot it built, and asks whether damaged-order cases should go to email or to a live chat.

A is the more confident proposal. B is the one you can test: its answer is already a line in your acceptance plan. That is the difference to look for in every proposal, whatever its position or badge.

A shortlist scorecard you can copy

Four or more out of six goes on the shortlist. Keep the shortlist to three; more interviews rarely change the answer and cost you a week.

Three questions for the shortlist

Send the same three to each shortlisted freelancer, in writing, so you can compare answers side by side:

1. Here are five real questions our customers ask [or: five real inputs this workflow will get]. How would your build handle each, including the one it should refuse or pass to a person?
2. How will you test it before delivery, and what test would you suggest I run before I approve the final payment?
3. After delivery, when something changes (our documents, a connected tool, a login), how do fixes work, and at what rate?

The first question is the most revealing for a chatbot. A good answer goes case by case and is honest about the one it cannot handle. The chatbot testing guide explains why the refusals and hand-overs matter as much as the answers. For an automation, ask about the failure path instead; the acceptance checklist has the tests worth naming.

The second question shows whether testing is part of how they work or something you will have to ask for. The third opens the maintenance conversation while you still have something to offer; the maintenance guide has the six lines to agree.

Before you hire: one more check

Write the answers into the contract terms or the first milestone description, not only the chat. “Handles the five sample questions as agreed, refuses the out-of-scope one, hands over to a person when asked” is a milestone you can test. The chatbot brief guide lists what to hand over at the start so those answers stay true during the build, and the AI chatbot job page has acceptance lines you can reuse.

Where it goes wrong anyway

• Hiring the first good proposal. Early proposals are often the fastest, not the closest fit. Wait until the post has collected answers, then score them together.
• Interviewing ten people. You will remember the most confident, not the most relevant. Three, same questions, in writing.
• Treating a badge as a reference. A high score tells you past contracts ended well. Your own three questions tell you whether this one will.
• Skipping the failure question because the demo looked good. Demos show the questions the builder expected. Your five real questions show the rest.

Wrap-up

A pile of proposals is mostly noise. Score each on whether it answers your job, shows close past work, asks about failure and fits your project’s size. Shortlist three, send all three the same questions, and hire the one whose answers you could turn into a test. If you have not posted yet, post the job on Upwork first, then score what comes back.